drivers/md/raid5.c

   1 /*
   2  * raid5.c : Multiple Devices driver for Linux
   3  *         Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   4  *         Copyright (C) 1999, 2000 Ingo Molnar
   5  *         Copyright (C) 2002, 2003 H. Peter Anvin
   6  *
   7  * RAID-4/5/6 management functions.
   8  * Thanks to Penguin Computing for making the RAID-6 development possible
   9  * by donating a test server!
  10  *
  11  * This program is free software; you can redistribute it and/or modify
  12  * it under the terms of the GNU General Public License as published by
  13  * the Free Software Foundation; either version 2, or (at your option)
  14  * any later version.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * (for example /usr/src/linux/COPYING); if not, write to the Free
  18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  19  */
  20
  21 /*
  22  * BITMAP UNPLUGGING:
  23  *
  24  * The sequencing for updating the bitmap reliably is a little
  25  * subtle (and I got it wrong the first time) so it deserves some
  26  * explanation.
  27  *
  28  * We group bitmap updates into batches.  Each batch has a number.
  29  * We may write out several batches at once, but that isn't very important.
  30  * conf->seq_write is the number of the last batch successfully written.
  31  * conf->seq_flush is the number of the last batch that was closed to
  32  *    new additions.
  33  * When we discover that we will need to write to any block in a stripe
  34  * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
  35  * the number of the batch it will be in. This is seq_flush+1.
  36  * When we are ready to do a write, if that batch hasn't been written yet,
  37  *   we plug the array and queue the stripe for later.
  38  * When an unplug happens, we increment bm_flush, thus closing the current
  39  *   batch.
  40  * When we notice that bm_flush > bm_write, we write out all pending updates
  41  * to the bitmap, and advance bm_write to where bm_flush was.
  42  * This may occasionally write a bit out twice, but is sure never to
  43  * miss any bits.
  44  */
  45
  46 #include <linux/blkdev.h>
  47 #include <linux/kthread.h>
  48 #include <linux/raid/pq.h>
  49 #include <linux/async_tx.h>
  50 #include <linux/module.h>
  51 #include <linux/async.h>
  52 #include <linux/seq_file.h>
  53 #include <linux/cpu.h>
  54 #include <linux/slab.h>
  55 #include <linux/ratelimit.h>
  56 #include <linux/nodemask.h>
  57 #include <linux/flex_array.h>
  58 #include <linux/sched/signal.h>
  59
  60 #include <trace/events/block.h>
  61 #include <linux/list_sort.h>
  62
  63 #include "md.h"
  64 #include "raid5.h"
  65 #include "raid0.h"
  66 #include "bitmap.h"
  67 #include "raid5-log.h"
  68
  69 #define UNSUPPORTED_MDDEV_FLAGS (1L << MD_FAILFAST_SUPPORTED)
  70
  71 #define cpu_to_group(cpu) cpu_to_node(cpu)
  72 #define ANY_GROUP NUMA_NO_NODE
  73
  74 static bool devices_handle_discard_safely = false;
  75 module_param(devices_handle_discard_safely, bool, 0644);
  76 MODULE_PARM_DESC(devices_handle_discard_safely,
  77                  "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
  78 static struct workqueue_struct *raid5_wq;
  79
  80 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
  81 {
  82         int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
  83         return &conf->stripe_hashtbl[hash];
  84 }
  85
  86 static inline int stripe_hash_locks_hash(sector_t sect)
  87 {
  88         return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
  89 }
  90
  91 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
  92 {
  93         spin_lock_irq(conf->hash_locks + hash);
  94         spin_lock(&conf->device_lock);
  95 }
  96
  97 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
  98 {
  99         spin_unlock(&conf->device_lock);
 100         spin_unlock_irq(conf->hash_locks + hash);
 101 }
 102
 103 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
 104 {
 105         int i;
 106         spin_lock_irq(conf->hash_locks);
 107         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
 108                 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
 109         spin_lock(&conf->device_lock);
 110 }
 111
 112 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
 113 {
 114         int i;
 115         spin_unlock(&conf->device_lock);
 116         for (i = NR_STRIPE_HASH_LOCKS - 1; i; i--)
 117                 spin_unlock(conf->hash_locks + i);
 118         spin_unlock_irq(conf->hash_locks);
 119 }
 120
 121 /* Find first data disk in a raid6 stripe */
 122 static inline int raid6_d0(struct stripe_head *sh)
 123 {
 124         if (sh->ddf_layout)
 125                 /* ddf always start from first device */
 126                 return 0;
 127         /* md starts just after Q block */
 128         if (sh->qd_idx == sh->disks - 1)
 129                 return 0;
 130         else
 131                 return sh->qd_idx + 1;
 132 }
 133 static inline int raid6_next_disk(int disk, int raid_disks)
 134 {
 135         disk++;
 136         return (disk < raid_disks) ? disk : 0;
 137 }
 138
 139 /* When walking through the disks in a raid5, starting at raid6_d0,
 140  * We need to map each disk to a 'slot', where the data disks are slot
 141  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
 142  * is raid_disks-1.  This help does that mapping.
 143  */
 144 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
 145                              int *count, int syndrome_disks)
 146 {
 147         int slot = *count;
 148
 149         if (sh->ddf_layout)
 150                 (*count)++;
 151         if (idx == sh->pd_idx)
 152                 return syndrome_disks;
 153         if (idx == sh->qd_idx)
 154                 return syndrome_disks + 1;
 155         if (!sh->ddf_layout)
 156                 (*count)++;
 157         return slot;
 158 }
 159
 160 static void print_raid5_conf (struct r5conf *conf);
 161
 162 static int stripe_operations_active(struct stripe_head *sh)
 163 {
 164         return sh->check_state || sh->reconstruct_state ||
 165                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
 166                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
 167 }
 168
 169 static bool stripe_is_lowprio(struct stripe_head *sh)
 170 {
 171         return (test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) ||
 172                 test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) &&
 173                !test_bit(STRIPE_R5C_CACHING, &sh->state);
 174 }
 175
 176 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
 177 {
 178         struct r5conf *conf = sh->raid_conf;
 179         struct r5worker_group *group;
 180         int thread_cnt;
 181         int i, cpu = sh->cpu;
 182
 183         if (!cpu_online(cpu)) {
 184                 cpu = cpumask_any(cpu_online_mask);
 185                 sh->cpu = cpu;
 186         }
 187
 188         if (list_empty(&sh->lru)) {
 189                 struct r5worker_group *group;
 190                 group = conf->worker_groups + cpu_to_group(cpu);
 191                 if (stripe_is_lowprio(sh))
 192                         list_add_tail(&sh->lru, &group->loprio_list);
 193                 else
 194                         list_add_tail(&sh->lru, &group->handle_list);
 195                 group->stripes_cnt++;
 196                 sh->group = group;
 197         }
 198
 199         if (conf->worker_cnt_per_group == 0) {
 200                 md_wakeup_thread(conf->mddev->thread);
 201                 return;
 202         }
 203
 204         group = conf->worker_groups + cpu_to_group(sh->cpu);
 205
 206         group->workers[0].working = true;
 207         /* at least one worker should run to avoid race */
 208         queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
 209
 210         thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
 211         /* wakeup more workers */
 212         for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
 213                 if (group->workers[i].working == false) {
 214                         group->workers[i].working = true;
 215                         queue_work_on(sh->cpu, raid5_wq,
 216                                       &group->workers[i].work);
 217                         thread_cnt--;
 218                 }
 219         }
 220 }
 221
 222 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
 223                               struct list_head *temp_inactive_list)
 224 {
 225         int i;
 226         int injournal = 0;      /* number of date pages with R5_InJournal */
 227
 228         BUG_ON(!list_empty(&sh->lru));
 229         BUG_ON(atomic_read(&conf->active_stripes)==0);
 230
 231         if (r5c_is_writeback(conf->log))
 232                 for (i = sh->disks; i--; )
 233                         if (test_bit(R5_InJournal, &sh->dev[i].flags))
 234                                 injournal++;
 235         /*
 236          * In the following cases, the stripe cannot be released to cached
 237          * lists. Therefore, we make the stripe write out and set
 238          * STRIPE_HANDLE:
 239          *   1. when quiesce in r5c write back;
 240          *   2. when resync is requested fot the stripe.
 241          */
 242         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) ||
 243             (conf->quiesce && r5c_is_writeback(conf->log) &&
 244              !test_bit(STRIPE_HANDLE, &sh->state) && injournal != 0)) {
 245                 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
 246                         r5c_make_stripe_write_out(sh);
 247                 set_bit(STRIPE_HANDLE, &sh->state);
 248         }
 249
 250         if (test_bit(STRIPE_HANDLE, &sh->state)) {
 251                 if (test_bit(STRIPE_DELAYED, &sh->state) &&
 252                     !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 253                         list_add_tail(&sh->lru, &conf->delayed_list);
 254                 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
 255                            sh->bm_seq - conf->seq_write > 0)
 256                         list_add_tail(&sh->lru, &conf->bitmap_list);
 257                 else {
 258                         clear_bit(STRIPE_DELAYED, &sh->state);
 259                         clear_bit(STRIPE_BIT_DELAY, &sh->state);
 260                         if (conf->worker_cnt_per_group == 0) {
 261                                 if (stripe_is_lowprio(sh))
 262                                         list_add_tail(&sh->lru,
 263                                                         &conf->loprio_list);
 264                                 else
 265                                         list_add_tail(&sh->lru,
 266                                                         &conf->handle_list);
 267                         } else {
 268                                 raid5_wakeup_stripe_thread(sh);
 269                                 return;
 270                         }
 271                 }
 272                 md_wakeup_thread(conf->mddev->thread);
 273         } else {
 274                 BUG_ON(stripe_operations_active(sh));
 275                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 276                         if (atomic_dec_return(&conf->preread_active_stripes)
 277                             < IO_THRESHOLD)
 278                                 md_wakeup_thread(conf->mddev->thread);
 279                 atomic_dec(&conf->active_stripes);
 280                 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
 281                         if (!r5c_is_writeback(conf->log))
 282                                 list_add_tail(&sh->lru, temp_inactive_list);
 283                         else {
 284                                 WARN_ON(test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags));
 285                                 if (injournal == 0)
 286                                         list_add_tail(&sh->lru, temp_inactive_list);
 287                                 else if (injournal == conf->raid_disks - conf->max_degraded) {
 288                                         /* full stripe */
 289                                         if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE, &sh->state))
 290                                                 atomic_inc(&conf->r5c_cached_full_stripes);
 291                                         if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
 292                                                 atomic_dec(&conf->r5c_cached_partial_stripes);
 293                                         list_add_tail(&sh->lru, &conf->r5c_full_stripe_list);
 294                                         r5c_check_cached_full_stripe(conf);
 295                                 } else
 296                                         /*
 297                                          * STRIPE_R5C_PARTIAL_STRIPE is set in
 298                                          * r5c_try_caching_write(). No need to
 299                                          * set it again.
 300                                          */
 301                                         list_add_tail(&sh->lru, &conf->r5c_partial_stripe_list);
 302                         }
 303                 }
 304         }
 305 }
 306
 307 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
 308                              struct list_head *temp_inactive_list)
 309 {
 310         if (atomic_dec_and_test(&sh->count))
 311                 do_release_stripe(conf, sh, temp_inactive_list);
 312 }
 313
 314 /*
 315  * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
 316  *
 317  * Be careful: Only one task can add/delete stripes from temp_inactive_list at
 318  * given time. Adding stripes only takes device lock, while deleting stripes
 319  * only takes hash lock.
 320  */
 321 static void release_inactive_stripe_list(struct r5conf *conf,
 322                                          struct list_head *temp_inactive_list,
 323                                          int hash)
 324 {
 325         int size;
 326         bool do_wakeup = false;
 327         unsigned long flags;
 328
 329         if (hash == NR_STRIPE_HASH_LOCKS) {
 330                 size = NR_STRIPE_HASH_LOCKS;
 331                 hash = NR_STRIPE_HASH_LOCKS - 1;
 332         } else
 333                 size = 1;
 334         while (size) {
 335                 struct list_head *list = &temp_inactive_list[size - 1];
 336
 337                 /*
 338                  * We don't hold any lock here yet, raid5_get_active_stripe() might
 339                  * remove stripes from the list
 340                  */
 341                 if (!list_empty_careful(list)) {
 342                         spin_lock_irqsave(conf->hash_locks + hash, flags);
 343                         if (list_empty(conf->inactive_list + hash) &&
 344                             !list_empty(list))
 345                                 atomic_dec(&conf->empty_inactive_list_nr);
 346                         list_splice_tail_init(list, conf->inactive_list + hash);
 347                         do_wakeup = true;
 348                         spin_unlock_irqrestore(conf->hash_locks + hash, flags);
 349                 }
 350                 size--;
 351                 hash--;
 352         }
 353
 354         if (do_wakeup) {
 355                 wake_up(&conf->wait_for_stripe);
 356                 if (atomic_read(&conf->active_stripes) == 0)
 357                         wake_up(&conf->wait_for_quiescent);
 358                 if (conf->retry_read_aligned)
 359                         md_wakeup_thread(conf->mddev->thread);
 360         }
 361 }
 362
 363 /* should hold conf->device_lock already */
 364 static int release_stripe_list(struct r5conf *conf,
 365                                struct list_head *temp_inactive_list)
 366 {
 367         struct stripe_head *sh, *t;
 368         int count = 0;
 369         struct llist_node *head;
 370
 371         head = llist_del_all(&conf->released_stripes);
 372         head = llist_reverse_order(head);
 373         llist_for_each_entry_safe(sh, t, head, release_list) {
 374                 int hash;
 375
 376                 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
 377                 smp_mb();
 378                 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
 379                 /*
 380                  * Don't worry the bit is set here, because if the bit is set
 381                  * again, the count is always > 1. This is true for
 382                  * STRIPE_ON_UNPLUG_LIST bit too.
 383                  */
 384                 hash = sh->hash_lock_index;
 385                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
 386                 count++;
 387         }
 388
 389         return count;
 390 }
 391
 392 void raid5_release_stripe(struct stripe_head *sh)
 393 {
 394         struct r5conf *conf = sh->raid_conf;
 395         unsigned long flags;
 396         struct list_head list;
 397         int hash;
 398         bool wakeup;
 399
 400         /* Avoid release_list until the last reference.
 401          */
 402         if (atomic_add_unless(&sh->count, -1, 1))
 403                 return;
 404
 405         if (unlikely(!conf->mddev->thread) ||
 406                 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
 407                 goto slow_path;
 408         wakeup = llist_add(&sh->release_list, &conf->released_stripes);
 409         if (wakeup)
 410                 md_wakeup_thread(conf->mddev->thread);
 411         return;
 412 slow_path:
 413         local_irq_save(flags);
 414         /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
 415         if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
 416                 INIT_LIST_HEAD(&list);
 417                 hash = sh->hash_lock_index;
 418                 do_release_stripe(conf, sh, &list);
 419                 spin_unlock(&conf->device_lock);
 420                 release_inactive_stripe_list(conf, &list, hash);
 421         }
 422         local_irq_restore(flags);
 423 }
 424
 425 static inline void remove_hash(struct stripe_head *sh)
 426 {
 427         pr_debug("remove_hash(), stripe %llu\n",
 428                 (unsigned long long)sh->sector);
 429
 430         hlist_del_init(&sh->hash);
 431 }
 432
 433 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
 434 {
 435         struct hlist_head *hp = stripe_hash(conf, sh->sector);
 436
 437         pr_debug("insert_hash(), stripe %llu\n",
 438                 (unsigned long long)sh->sector);
 439
 440         hlist_add_head(&sh->hash, hp);
 441 }
 442
 443 /* find an idle stripe, make sure it is unhashed, and return it. */
 444 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
 445 {
 446         struct stripe_head *sh = NULL;
 447         struct list_head *first;
 448
 449         if (list_empty(conf->inactive_list + hash))
 450                 goto out;
 451         first = (conf->inactive_list + hash)->next;
 452         sh = list_entry(first, struct stripe_head, lru);
 453         list_del_init(first);
 454         remove_hash(sh);
 455         atomic_inc(&conf->active_stripes);
 456         BUG_ON(hash != sh->hash_lock_index);
 457         if (list_empty(conf->inactive_list + hash))
 458                 atomic_inc(&conf->empty_inactive_list_nr);
 459 out:
 460         return sh;
 461 }
 462
 463 static void shrink_buffers(struct stripe_head *sh)
 464 {
 465         struct page *p;
 466         int i;
 467         int num = sh->raid_conf->pool_size;
 468
 469         for (i = 0; i < num ; i++) {
 470                 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
 471                 p = sh->dev[i].page;
 472                 if (!p)
 473                         continue;
 474                 sh->dev[i].page = NULL;
 475                 put_page(p);
 476         }
 477 }
 478
 479 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
 480 {
 481         int i;
 482         int num = sh->raid_conf->pool_size;
 483
 484         for (i = 0; i < num; i++) {
 485                 struct page *page;
 486
 487                 if (!(page = alloc_page(gfp))) {
 488                         return 1;
 489                 }
 490                 sh->dev[i].page = page;
 491                 sh->dev[i].orig_page = page;
 492         }
 493
 494         return 0;
 495 }
 496
 497 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
 498                             struct stripe_head *sh);
 499
 500 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
 501 {
 502         struct r5conf *conf = sh->raid_conf;
 503         int i, seq;
 504
 505         BUG_ON(atomic_read(&sh->count) != 0);
 506         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 507         BUG_ON(stripe_operations_active(sh));
 508         BUG_ON(sh->batch_head);
 509
 510         pr_debug("init_stripe called, stripe %llu\n",
 511                 (unsigned long long)sector);
 512 retry:
 513         seq = read_seqcount_begin(&conf->gen_lock);
 514         sh->generation = conf->generation - previous;
 515         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
 516         sh->sector = sector;
 517         stripe_set_idx(sector, conf, previous, sh);
 518         sh->state = 0;
 519
 520         for (i = sh->disks; i--; ) {
 521                 struct r5dev *dev = &sh->dev[i];
 522
 523                 if (dev->toread || dev->read || dev->towrite || dev->written ||
 524                     test_bit(R5_LOCKED, &dev->flags)) {
 525                         pr_err("sector=%llx i=%d %p %p %p %p %d\n",
 526                                (unsigned long long)sh->sector, i, dev->toread,
 527                                dev->read, dev->towrite, dev->written,
 528                                test_bit(R5_LOCKED, &dev->flags));
 529                         WARN_ON(1);
 530                 }
 531                 dev->flags = 0;
 532                 dev->sector = raid5_compute_blocknr(sh, i, previous);
 533         }
 534         if (read_seqcount_retry(&conf->gen_lock, seq))
 535                 goto retry;
 536         sh->overwrite_disks = 0;
 537         insert_hash(conf, sh);
 538         sh->cpu = smp_processor_id();
 539         set_bit(STRIPE_BATCH_READY, &sh->state);
 540 }
 541
 542 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
 543                                          short generation)
 544 {
 545         struct stripe_head *sh;
 546
 547         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
 548         hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
 549                 if (sh->sector == sector && sh->generation == generation)
 550                         return sh;
 551         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
 552         return NULL;
 553 }
 554
 555 /*
 556  * Need to check if array has failed when deciding whether to:
 557  *  - start an array
 558  *  - remove non-faulty devices
 559  *  - add a spare
 560  *  - allow a reshape
 561  * This determination is simple when no reshape is happening.
 562  * However if there is a reshape, we need to carefully check
 563  * both the before and after sections.
 564  * This is because some failed devices may only affect one
 565  * of the two sections, and some non-in_sync devices may
 566  * be insync in the section most affected by failed devices.
 567  */
 568 int raid5_calc_degraded(struct r5conf *conf)
 569 {
 570         int degraded, degraded2;
 571         int i;
 572
 573         rcu_read_lock();
 574         degraded = 0;
 575         for (i = 0; i < conf->previous_raid_disks; i++) {
 576                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 577                 if (rdev && test_bit(Faulty, &rdev->flags))
 578                         rdev = rcu_dereference(conf->disks[i].replacement);
 579                 if (!rdev || test_bit(Faulty, &rdev->flags))
 580                         degraded++;
 581                 else if (test_bit(In_sync, &rdev->flags))
 582                         ;
 583                 else
 584                         /* not in-sync or faulty.
 585                          * If the reshape increases the number of devices,
 586                          * this is being recovered by the reshape, so
 587                          * this 'previous' section is not in_sync.
 588                          * If the number of devices is being reduced however,
 589                          * the device can only be part of the array if
 590                          * we are reverting a reshape, so this section will
 591                          * be in-sync.
 592                          */
 593                         if (conf->raid_disks >= conf->previous_raid_disks)
 594                                 degraded++;
 595         }
 596         rcu_read_unlock();
 597         if (conf->raid_disks == conf->previous_raid_disks)
 598                 return degraded;
 599         rcu_read_lock();
 600         degraded2 = 0;
 601         for (i = 0; i < conf->raid_disks; i++) {
 602                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 603                 if (rdev && test_bit(Faulty, &rdev->flags))
 604                         rdev = rcu_dereference(conf->disks[i].replacement);
 605                 if (!rdev || test_bit(Faulty, &rdev->flags))
 606                         degraded2++;
 607                 else if (test_bit(In_sync, &rdev->flags))
 608                         ;
 609                 else
 610                         /* not in-sync or faulty.
 611                          * If reshape increases the number of devices, this
 612                          * section has already been recovered, else it
 613                          * almost certainly hasn't.
 614                          */
 615                         if (conf->raid_disks <= conf->previous_raid_disks)
 616                                 degraded2++;
 617         }
 618         rcu_read_unlock();
 619         if (degraded2 > degraded)
 620                 return degraded2;
 621         return degraded;
 622 }
 623
 624 static int has_failed(struct r5conf *conf)
 625 {
 626         int degraded;
 627
 628         if (conf->mddev->reshape_position == MaxSector)
 629                 return conf->mddev->degraded > conf->max_degraded;
 630
 631         degraded = raid5_calc_degraded(conf);
 632         if (degraded > conf->max_degraded)
 633                 return 1;
 634         return 0;
 635 }
 636
 637 struct stripe_head *
 638 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
 639                         int previous, int noblock, int noquiesce)
 640 {
 641         struct stripe_head *sh;
 642         int hash = stripe_hash_locks_hash(sector);
 643         int inc_empty_inactive_list_flag;
 644
 645         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 646
 647         spin_lock_irq(conf->hash_locks + hash);
 648
 649         do {
 650                 wait_event_lock_irq(conf->wait_for_quiescent,
 651                                     conf->quiesce == 0 || noquiesce,
 652                                     *(conf->hash_locks + hash));
 653                 sh = __find_stripe(conf, sector, conf->generation - previous);
 654                 if (!sh) {
 655                         if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
 656                                 sh = get_free_stripe(conf, hash);
 657                                 if (!sh && !test_bit(R5_DID_ALLOC,
 658                                                      &conf->cache_state))
 659                                         set_bit(R5_ALLOC_MORE,
 660                                                 &conf->cache_state);
 661                         }
 662                         if (noblock && sh == NULL)
 663                                 break;
 664
 665                         r5c_check_stripe_cache_usage(conf);
 666                         if (!sh) {
 667                                 set_bit(R5_INACTIVE_BLOCKED,
 668                                         &conf->cache_state);
 669                                 r5l_wake_reclaim(conf->log, 0);
 670                                 wait_event_lock_irq(
 671                                         conf->wait_for_stripe,
 672                                         !list_empty(conf->inactive_list + hash) &&
 673                                         (atomic_read(&conf->active_stripes)
 674                                          < (conf->max_nr_stripes * 3 / 4)
 675                                          || !test_bit(R5_INACTIVE_BLOCKED,
 676                                                       &conf->cache_state)),
 677                                         *(conf->hash_locks + hash));
 678                                 clear_bit(R5_INACTIVE_BLOCKED,
 679                                           &conf->cache_state);
 680                         } else {
 681                                 init_stripe(sh, sector, previous);
 682                                 atomic_inc(&sh->count);
 683                         }
 684                 } else if (!atomic_inc_not_zero(&sh->count)) {
 685                         spin_lock(&conf->device_lock);
 686                         if (!atomic_read(&sh->count)) {
 687                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 688                                         atomic_inc(&conf->active_stripes);
 689                                 BUG_ON(list_empty(&sh->lru) &&
 690                                        !test_bit(STRIPE_EXPANDING, &sh->state));
 691                                 inc_empty_inactive_list_flag = 0;
 692                                 if (!list_empty(conf->inactive_list + hash))
 693                                         inc_empty_inactive_list_flag = 1;
 694                                 list_del_init(&sh->lru);
 695                                 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
 696                                         atomic_inc(&conf->empty_inactive_list_nr);
 697                                 if (sh->group) {
 698                                         sh->group->stripes_cnt--;
 699                                         sh->group = NULL;
 700                                 }
 701                         }
 702                         atomic_inc(&sh->count);
 703                         spin_unlock(&conf->device_lock);
 704                 }
 705         } while (sh == NULL);
 706
 707         spin_unlock_irq(conf->hash_locks + hash);
 708         return sh;
 709 }
 710
 711 static bool is_full_stripe_write(struct stripe_head *sh)
 712 {
 713         BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
 714         return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
 715 }
 716
 717 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 718 {
 719         if (sh1 > sh2) {
 720                 spin_lock_irq(&sh2->stripe_lock);
 721                 spin_lock_nested(&sh1->stripe_lock, 1);
 722         } else {
 723                 spin_lock_irq(&sh1->stripe_lock);
 724                 spin_lock_nested(&sh2->stripe_lock, 1);
 725         }
 726 }
 727
 728 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 729 {
 730         spin_unlock(&sh1->stripe_lock);
 731         spin_unlock_irq(&sh2->stripe_lock);
 732 }
 733
 734 /* Only freshly new full stripe normal write stripe can be added to a batch list */
 735 static bool stripe_can_batch(struct stripe_head *sh)
 736 {
 737         struct r5conf *conf = sh->raid_conf;
 738
 739         if (conf->log || raid5_has_ppl(conf))
 740                 return false;
 741         return test_bit(STRIPE_BATCH_READY, &sh->state) &&
 742                 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
 743                 is_full_stripe_write(sh);
 744 }
 745
 746 /* we only do back search */
 747 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
 748 {
 749         struct stripe_head *head;
 750         sector_t head_sector, tmp_sec;
 751         int hash;
 752         int dd_idx;
 753         int inc_empty_inactive_list_flag;
 754
 755         /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
 756         tmp_sec = sh->sector;
 757         if (!sector_div(tmp_sec, conf->chunk_sectors))
 758                 return;
 759         head_sector = sh->sector - STRIPE_SECTORS;
 760
 761         hash = stripe_hash_locks_hash(head_sector);
 762         spin_lock_irq(conf->hash_locks + hash);
 763         head = __find_stripe(conf, head_sector, conf->generation);
 764         if (head && !atomic_inc_not_zero(&head->count)) {
 765                 spin_lock(&conf->device_lock);
 766                 if (!atomic_read(&head->count)) {
 767                         if (!test_bit(STRIPE_HANDLE, &head->state))
 768                                 atomic_inc(&conf->active_stripes);
 769                         BUG_ON(list_empty(&head->lru) &&
 770                                !test_bit(STRIPE_EXPANDING, &head->state));
 771                         inc_empty_inactive_list_flag = 0;
 772                         if (!list_empty(conf->inactive_list + hash))
 773                                 inc_empty_inactive_list_flag = 1;
 774                         list_del_init(&head->lru);
 775                         if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
 776                                 atomic_inc(&conf->empty_inactive_list_nr);
 777                         if (head->group) {
 778                                 head->group->stripes_cnt--;
 779                                 head->group = NULL;
 780                         }
 781                 }
 782                 atomic_inc(&head->count);
 783                 spin_unlock(&conf->device_lock);
 784         }
 785         spin_unlock_irq(conf->hash_locks + hash);
 786
 787         if (!head)
 788                 return;
 789         if (!stripe_can_batch(head))
 790                 goto out;
 791
 792         lock_two_stripes(head, sh);
 793         /* clear_batch_ready clear the flag */
 794         if (!stripe_can_batch(head) || !stripe_can_batch(sh))
 795                 goto unlock_out;
 796
 797         if (sh->batch_head)
 798                 goto unlock_out;
 799
 800         dd_idx = 0;
 801         while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
 802                 dd_idx++;
 803         if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
 804             bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
 805                 goto unlock_out;
 806
 807         if (head->batch_head) {
 808                 spin_lock(&head->batch_head->batch_lock);
 809                 /* This batch list is already running */
 810                 if (!stripe_can_batch(head)) {
 811                         spin_unlock(&head->batch_head->batch_lock);
 812                         goto unlock_out;
 813                 }
 814                 /*
 815                  * We must assign batch_head of this stripe within the
 816                  * batch_lock, otherwise clear_batch_ready of batch head
 817                  * stripe could clear BATCH_READY bit of this stripe and
 818                  * this stripe->batch_head doesn't get assigned, which
 819                  * could confuse clear_batch_ready for this stripe
 820                  */
 821                 sh->batch_head = head->batch_head;
 822
 823                 /*
 824                  * at this point, head's BATCH_READY could be cleared, but we
 825                  * can still add the stripe to batch list
 826                  */
 827                 list_add(&sh->batch_list, &head->batch_list);
 828                 spin_unlock(&head->batch_head->batch_lock);
 829         } else {
 830                 head->batch_head = head;
 831                 sh->batch_head = head->batch_head;
 832                 spin_lock(&head->batch_lock);
 833                 list_add_tail(&sh->batch_list, &head->batch_list);
 834                 spin_unlock(&head->batch_lock);
 835         }
 836
 837         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 838                 if (atomic_dec_return(&conf->preread_active_stripes)
 839                     < IO_THRESHOLD)
 840                         md_wakeup_thread(conf->mddev->thread);
 841
 842         if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
 843                 int seq = sh->bm_seq;
 844                 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
 845                     sh->batch_head->bm_seq > seq)
 846                         seq = sh->batch_head->bm_seq;
 847                 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
 848                 sh->batch_head->bm_seq = seq;
 849         }
 850
 851         atomic_inc(&sh->count);
 852 unlock_out:
 853         unlock_two_stripes(head, sh);
 854 out:
 855         raid5_release_stripe(head);
 856 }
 857
 858 /* Determine if 'data_offset' or 'new_data_offset' should be used
 859  * in this stripe_head.
 860  */
 861 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
 862 {
 863         sector_t progress = conf->reshape_progress;
 864         /* Need a memory barrier to make sure we see the value
 865          * of conf->generation, or ->data_offset that was set before
 866          * reshape_progress was updated.
 867          */
 868         smp_rmb();
 869         if (progress == MaxSector)
 870                 return 0;
 871         if (sh->generation == conf->generation - 1)
 872                 return 0;
 873         /* We are in a reshape, and this is a new-generation stripe,
 874          * so use new_data_offset.
 875          */
 876         return 1;
 877 }
 878
 879 static void dispatch_bio_list(struct bio_list *tmp)
 880 {
 881         struct bio *bio;
 882
 883         while ((bio = bio_list_pop(tmp)))
 884                 generic_make_request(bio);
 885 }
 886
 887 static int cmp_stripe(void *priv, struct list_head *a, struct list_head *b)
 888 {
 889         const struct r5pending_data *da = list_entry(a,
 890                                 struct r5pending_data, sibling);
 891         const struct r5pending_data *db = list_entry(b,
 892                                 struct r5pending_data, sibling);
 893         if (da->sector > db->sector)
 894                 return 1;
 895         if (da->sector < db->sector)
 896                 return -1;
 897         return 0;
 898 }
 899
 900 static void dispatch_defer_bios(struct r5conf *conf, int target,
 901                                 struct bio_list *list)
 902 {
 903         struct r5pending_data *data;
 904         struct list_head *first, *next = NULL;
 905         int cnt = 0;
 906
 907         if (conf->pending_data_cnt == 0)
 908                 return;
 909
 910         list_sort(NULL, &conf->pending_list, cmp_stripe);
 911
 912         first = conf->pending_list.next;
 913
 914         /* temporarily move the head */
 915         if (conf->next_pending_data)
 916                 list_move_tail(&conf->pending_list,
 917                                 &conf->next_pending_data->sibling);
 918
 919         while (!list_empty(&conf->pending_list)) {
 920                 data = list_first_entry(&conf->pending_list,
 921                         struct r5pending_data, sibling);
 922                 if (&data->sibling == first)
 923                         first = data->sibling.next;
 924                 next = data->sibling.next;
 925
 926                 bio_list_merge(list, &data->bios);
 927                 list_move(&data->sibling, &conf->free_list);
 928                 cnt++;
 929                 if (cnt >= target)
 930                         break;
 931         }
 932         conf->pending_data_cnt -= cnt;
 933         BUG_ON(conf->pending_data_cnt < 0 || cnt < target);
 934
 935         if (next != &conf->pending_list)
 936                 conf->next_pending_data = list_entry(next,
 937                                 struct r5pending_data, sibling);
 938         else
 939                 conf->next_pending_data = NULL;
 940         /* list isn't empty */
 941         if (first != &conf->pending_list)
 942                 list_move_tail(&conf->pending_list, first);
 943 }
 944
 945 static void flush_deferred_bios(struct r5conf *conf)
 946 {
 947         struct bio_list tmp = BIO_EMPTY_LIST;
 948
 949         if (conf->pending_data_cnt == 0)
 950                 return;
 951
 952         spin_lock(&conf->pending_bios_lock);
 953         dispatch_defer_bios(conf, conf->pending_data_cnt, &tmp);
 954         BUG_ON(conf->pending_data_cnt != 0);
 955         spin_unlock(&conf->pending_bios_lock);
 956
 957         dispatch_bio_list(&tmp);
 958 }
 959
 960 static void defer_issue_bios(struct r5conf *conf, sector_t sector,
 961                                 struct bio_list *bios)
 962 {
 963         struct bio_list tmp = BIO_EMPTY_LIST;
 964         struct r5pending_data *ent;
 965
 966         spin_lock(&conf->pending_bios_lock);
 967         ent = list_first_entry(&conf->free_list, struct r5pending_data,
 968                                                         sibling);
 969         list_move_tail(&ent->sibling, &conf->pending_list);
 970         ent->sector = sector;
 971         bio_list_init(&ent->bios);
 972         bio_list_merge(&ent->bios, bios);
 973         conf->pending_data_cnt++;
 974         if (conf->pending_data_cnt >= PENDING_IO_MAX)
 975                 dispatch_defer_bios(conf, PENDING_IO_ONE_FLUSH, &tmp);
 976
 977         spin_unlock(&conf->pending_bios_lock);
 978
 979         dispatch_bio_list(&tmp);
 980 }
 981
 982 static void
 983 raid5_end_read_request(struct bio *bi);
 984 static void
 985 raid5_end_write_request(struct bio *bi);
 986
 987 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
 988 {
 989         struct r5conf *conf = sh->raid_conf;
 990         int i, disks = sh->disks;
 991         struct stripe_head *head_sh = sh;
 992         struct bio_list pending_bios = BIO_EMPTY_LIST;
 993         bool should_defer;
 994
 995         might_sleep();
 996
 997         if (log_stripe(sh, s) == 0)
 998                 return;
 999
1000         should_defer = conf->batch_bio_dispatch && conf->group_cnt;
1001
1002         for (i = disks; i--; ) {
1003                 int op, op_flags = 0;
1004                 int replace_only = 0;
1005                 struct bio *bi, *rbi;
1006                 struct md_rdev *rdev, *rrdev = NULL;
1007
1008                 sh = head_sh;
1009                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
1010                         op = REQ_OP_WRITE;
1011                         if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
1012                                 op_flags = REQ_FUA;
1013                         if (test_bit(R5_Discard, &sh->dev[i].flags))
1014                                 op = REQ_OP_DISCARD;
1015                 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1016                         op = REQ_OP_READ;
1017                 else if (test_and_clear_bit(R5_WantReplace,
1018                                             &sh->dev[i].flags)) {
1019                         op = REQ_OP_WRITE;
1020                         replace_only = 1;
1021                 } else
1022                         continue;
1023                 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
1024                         op_flags |= REQ_SYNC;
1025
1026 again:
1027                 bi = &sh->dev[i].req;
1028                 rbi = &sh->dev[i].rreq; /* For writing to replacement */
1029
1030                 rcu_read_lock();
1031                 rrdev = rcu_dereference(conf->disks[i].replacement);
1032                 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
1033                 rdev = rcu_dereference(conf->disks[i].rdev);
1034                 if (!rdev) {
1035                         rdev = rrdev;
1036                         rrdev = NULL;
1037                 }
1038                 if (op_is_write(op)) {
1039                         if (replace_only)
1040                                 rdev = NULL;
1041                         if (rdev == rrdev)
1042                                 /* We raced and saw duplicates */
1043                                 rrdev = NULL;
1044                 } else {
1045                         if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
1046                                 rdev = rrdev;
1047                         rrdev = NULL;
1048                 }
1049
1050                 if (rdev && test_bit(Faulty, &rdev->flags))
1051                         rdev = NULL;
1052                 if (rdev)
1053                         atomic_inc(&rdev->nr_pending);
1054                 if (rrdev && test_bit(Faulty, &rrdev->flags))
1055                         rrdev = NULL;
1056                 if (rrdev)
1057                         atomic_inc(&rrdev->nr_pending);
1058                 rcu_read_unlock();
1059
1060                 /* We have already checked bad blocks for reads.  Now
1061                  * need to check for writes.  We never accept write errors
1062                  * on the replacement, so we don't to check rrdev.
1063                  */
1064                 while (op_is_write(op) && rdev &&
1065                        test_bit(WriteErrorSeen, &rdev->flags)) {
1066                         sector_t first_bad;
1067                         int bad_sectors;
1068                         int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
1069                                               &first_bad, &bad_sectors);
1070                         if (!bad)
1071                                 break;
1072
1073                         if (bad < 0) {
1074                                 set_bit(BlockedBadBlocks, &rdev->flags);
1075                                 if (!conf->mddev->external &&
1076                                     conf->mddev->sb_flags) {
1077                                         /* It is very unlikely, but we might
1078                                          * still need to write out the
1079                                          * bad block log - better give it
1080                                          * a chance*/
1081                                         md_check_recovery(conf->mddev);
1082                                 }
1083                                 /*
1084                                  * Because md_wait_for_blocked_rdev
1085                                  * will dec nr_pending, we must
1086                                  * increment it first.
1087                                  */
1088                                 atomic_inc(&rdev->nr_pending);
1089                                 md_wait_for_blocked_rdev(rdev, conf->mddev);
1090                         } else {
1091                                 /* Acknowledged bad block - skip the write */
1092                                 rdev_dec_pending(rdev, conf->mddev);
1093                                 rdev = NULL;
1094                         }
1095                 }
1096
1097                 if (rdev) {
1098                         if (s->syncing || s->expanding || s->expanded
1099                             || s->replacing)
1100                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1101
1102                         set_bit(STRIPE_IO_STARTED, &sh->state);
1103
1104                         bio_set_dev(bi, rdev->bdev);
1105                         bio_set_op_attrs(bi, op, op_flags);
1106                         bi->bi_end_io = op_is_write(op)
1107                                 ? raid5_end_write_request
1108                                 : raid5_end_read_request;
1109                         bi->bi_private = sh;
1110
1111                         pr_debug("%s: for %llu schedule op %d on disc %d\n",
1112                                 __func__, (unsigned long long)sh->sector,
1113                                 bi->bi_opf, i);
1114                         atomic_inc(&sh->count);
1115                         if (sh != head_sh)
1116                                 atomic_inc(&head_sh->count);
1117                         if (use_new_offset(conf, sh))
1118                                 bi->bi_iter.bi_sector = (sh->sector
1119                                                  + rdev->new_data_offset);
1120                         else
1121                                 bi->bi_iter.bi_sector = (sh->sector
1122                                                  + rdev->data_offset);
1123                         if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1124                                 bi->bi_opf |= REQ_NOMERGE;
1125
1126                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1127                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1128
1129                         if (!op_is_write(op) &&
1130                             test_bit(R5_InJournal, &sh->dev[i].flags))
1131                                 /*
1132                                  * issuing read for a page in journal, this
1133                                  * must be preparing for prexor in rmw; read
1134                                  * the data into orig_page
1135                                  */
1136                                 sh->dev[i].vec.bv_page = sh->dev[i].orig_page;
1137                         else
1138                                 sh->dev[i].vec.bv_page = sh->dev[i].page;
1139                         bi->bi_vcnt = 1;
1140                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1141                         bi->bi_io_vec[0].bv_offset = 0;
1142                         bi->bi_iter.bi_size = STRIPE_SIZE;
1143                         /*
1144                          * If this is discard request, set bi_vcnt 0. We don't
1145                          * want to confuse SCSI because SCSI will replace payload
1146                          */
1147                         if (op == REQ_OP_DISCARD)
1148                                 bi->bi_vcnt = 0;
1149                         if (rrdev)
1150                                 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1151
1152                         if (conf->mddev->gendisk)
1153                                 trace_block_bio_remap(bi->bi_disk->queue,
1154                                                       bi, disk_devt(conf->mddev->gendisk),
1155                                                       sh->dev[i].sector);
1156                         if (should_defer && op_is_write(op))
1157                                 bio_list_add(&pending_bios, bi);
1158                         else
1159                                 generic_make_request(bi);
1160                 }
1161                 if (rrdev) {
1162                         if (s->syncing || s->expanding || s->expanded
1163                             || s->replacing)
1164                                 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1165
1166                         set_bit(STRIPE_IO_STARTED, &sh->state);
1167
1168                         bio_set_dev(rbi, rrdev->bdev);
1169                         bio_set_op_attrs(rbi, op, op_flags);
1170                         BUG_ON(!op_is_write(op));
1171                         rbi->bi_end_io = raid5_end_write_request;
1172                         rbi->bi_private = sh;
1173
1174                         pr_debug("%s: for %llu schedule op %d on "
1175                                  "replacement disc %d\n",
1176                                 __func__, (unsigned long long)sh->sector,
1177                                 rbi->bi_opf, i);
1178                         atomic_inc(&sh->count);
1179                         if (sh != head_sh)
1180                                 atomic_inc(&head_sh->count);
1181                         if (use_new_offset(conf, sh))
1182                                 rbi->bi_iter.bi_sector = (sh->sector
1183                                                   + rrdev->new_data_offset);
1184                         else
1185                                 rbi->bi_iter.bi_sector = (sh->sector
1186                                                   + rrdev->data_offset);
1187                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1188                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1189                         sh->dev[i].rvec.bv_page = sh->dev[i].page;
1190                         rbi->bi_vcnt = 1;
1191                         rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1192                         rbi->bi_io_vec[0].bv_offset = 0;
1193                         rbi->bi_iter.bi_size = STRIPE_SIZE;
1194                         /*
1195                          * If this is discard request, set bi_vcnt 0. We don't
1196                          * want to confuse SCSI because SCSI will replace payload
1197                          */
1198                         if (op == REQ_OP_DISCARD)
1199                                 rbi->bi_vcnt = 0;
1200                         if (conf->mddev->gendisk)
1201                                 trace_block_bio_remap(rbi->bi_disk->queue,
1202                                                       rbi, disk_devt(conf->mddev->gendisk),
1203                                                       sh->dev[i].sector);
1204                         if (should_defer && op_is_write(op))
1205                                 bio_list_add(&pending_bios, rbi);
1206                         else
1207                                 generic_make_request(rbi);
1208                 }
1209                 if (!rdev && !rrdev) {
1210                         if (op_is_write(op))
1211                                 set_bit(STRIPE_DEGRADED, &sh->state);
1212                         pr_debug("skip op %d on disc %d for sector %llu\n",
1213                                 bi->bi_opf, i, (unsigned long long)sh->sector);
1214                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1215                         set_bit(STRIPE_HANDLE, &sh->state);
1216                 }
1217
1218                 if (!head_sh->batch_head)
1219                         continue;
1220                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1221                                       batch_list);
1222                 if (sh != head_sh)
1223                         goto again;
1224         }
1225
1226         if (should_defer && !bio_list_empty(&pending_bios))
1227                 defer_issue_bios(conf, head_sh->sector, &pending_bios);
1228 }
1229
1230 static struct dma_async_tx_descriptor *
1231 async_copy_data(int frombio, struct bio *bio, struct page **page,
1232         sector_t sector, struct dma_async_tx_descriptor *tx,
1233         struct stripe_head *sh, int no_skipcopy)
1234 {
1235         struct bio_vec bvl;
1236         struct bvec_iter iter;
1237         struct page *bio_page;
1238         int page_offset;
1239         struct async_submit_ctl submit;
1240         enum async_tx_flags flags = 0;
1241
1242         if (bio->bi_iter.bi_sector >= sector)
1243                 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1244         else
1245                 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1246
1247         if (frombio)
1248                 flags |= ASYNC_TX_FENCE;
1249         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1250
1251         bio_for_each_segment(bvl, bio, iter) {
1252                 int len = bvl.bv_len;
1253                 int clen;
1254                 int b_offset = 0;
1255
1256                 if (page_offset < 0) {
1257                         b_offset = -page_offset;
1258                         page_offset += b_offset;
1259                         len -= b_offset;
1260                 }
1261
1262                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1263                         clen = STRIPE_SIZE - page_offset;
1264                 else
1265                         clen = len;
1266
1267                 if (clen > 0) {
1268                         b_offset += bvl.bv_offset;
1269                         bio_page = bvl.bv_page;
1270                         if (frombio) {
1271                                 if (sh->raid_conf->skip_copy &&
1272                                     b_offset == 0 && page_offset == 0 &&
1273                                     clen == STRIPE_SIZE &&
1274                                     !no_skipcopy)
1275                                         *page = bio_page;
1276                                 else
1277                                         tx = async_memcpy(*page, bio_page, page_offset,
1278                                                   b_offset, clen, &submit);
1279                         } else
1280                                 tx = async_memcpy(bio_page, *page, b_offset,
1281                                                   page_offset, clen, &submit);
1282                 }
1283                 /* chain the operations */
1284                 submit.depend_tx = tx;
1285
1286                 if (clen < len) /* hit end of page */
1287                         break;
1288                 page_offset +=  len;
1289         }
1290
1291         return tx;
1292 }
1293
1294 static void ops_complete_biofill(void *stripe_head_ref)
1295 {
1296         struct stripe_head *sh = stripe_head_ref;
1297         int i;
1298
1299         pr_debug("%s: stripe %llu\n", __func__,
1300                 (unsigned long long)sh->sector);
1301
1302         /* clear completed biofills */
1303         for (i = sh->disks; i--; ) {
1304                 struct r5dev *dev = &sh->dev[i];
1305
1306                 /* acknowledge completion of a biofill operation */
1307                 /* and check if we need to reply to a read request,
1308                  * new R5_Wantfill requests are held off until
1309                  * !STRIPE_BIOFILL_RUN
1310                  */
1311                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1312                         struct bio *rbi, *rbi2;
1313
1314                         BUG_ON(!dev->read);
1315                         rbi = dev->read;
1316                         dev->read = NULL;
1317                         while (rbi && rbi->bi_iter.bi_sector <
1318                                 dev->sector + STRIPE_SECTORS) {
1319                                 rbi2 = r5_next_bio(rbi, dev->sector);
1320                                 bio_endio(rbi);
1321                                 rbi = rbi2;
1322                         }
1323                 }
1324         }
1325         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1326
1327         set_bit(STRIPE_HANDLE, &sh->state);
1328         raid5_release_stripe(sh);
1329 }
1330
1331 static void ops_run_biofill(struct stripe_head *sh)
1332 {
1333         struct dma_async_tx_descriptor *tx = NULL;
1334         struct async_submit_ctl submit;
1335         int i;
1336
1337         BUG_ON(sh->batch_head);
1338         pr_debug("%s: stripe %llu\n", __func__,
1339                 (unsigned long long)sh->sector);
1340
1341         for (i = sh->disks; i--; ) {
1342                 struct r5dev *dev = &sh->dev[i];
1343                 if (test_bit(R5_Wantfill, &dev->flags)) {
1344                         struct bio *rbi;
1345                         spin_lock_irq(&sh->stripe_lock);
1346                         dev->read = rbi = dev->toread;
1347                         dev->toread = NULL;
1348                         spin_unlock_irq(&sh->stripe_lock);
1349                         while (rbi && rbi->bi_iter.bi_sector <
1350                                 dev->sector + STRIPE_SECTORS) {
1351                                 tx = async_copy_data(0, rbi, &dev->page,
1352                                                      dev->sector, tx, sh, 0);
1353                                 rbi = r5_next_bio(rbi, dev->sector);
1354                         }
1355                 }
1356         }
1357
1358         atomic_inc(&sh->count);
1359         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1360         async_trigger_callback(&submit);
1361 }
1362
1363 static void mark_target_uptodate(struct stripe_head *sh, int target)
1364 {
1365         struct r5dev *tgt;
1366
1367         if (target < 0)
1368                 return;
1369
1370         tgt = &sh->dev[target];
1371         set_bit(R5_UPTODATE, &tgt->flags);
1372         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1373         clear_bit(R5_Wantcompute, &tgt->flags);
1374 }
1375
1376 static void ops_complete_compute(void *stripe_head_ref)
1377 {
1378         struct stripe_head *sh = stripe_head_ref;
1379
1380         pr_debug("%s: stripe %llu\n", __func__,
1381                 (unsigned long long)sh->sector);
1382
1383         /* mark the computed target(s) as uptodate */
1384         mark_target_uptodate(sh, sh->ops.target);
1385         mark_target_uptodate(sh, sh->ops.target2);
1386
1387         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1388         if (sh->check_state == check_state_compute_run)
1389                 sh->check_state = check_state_compute_result;
1390         set_bit(STRIPE_HANDLE, &sh->state);
1391         raid5_release_stripe(sh);
1392 }
1393
1394 /* return a pointer to the address conversion region of the scribble buffer */
1395 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1396                                  struct raid5_percpu *percpu, int i)
1397 {
1398         void *addr;
1399
1400         addr = flex_array_get(percpu->scribble, i);
1401         return addr + sizeof(struct page *) * (sh->disks + 2);
1402 }
1403
1404 /* return a pointer to the address conversion region of the scribble buffer */
1405 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1406 {
1407         void *addr;
1408
1409         addr = flex_array_get(percpu->scribble, i);
1410         return addr;
1411 }
1412
1413 static struct dma_async_tx_descriptor *
1414 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1415 {
1416         int disks = sh->disks;
1417         struct page **xor_srcs = to_addr_page(percpu, 0);
1418         int target = sh->ops.target;
1419         struct r5dev *tgt = &sh->dev[target];
1420         struct page *xor_dest = tgt->page;
1421         int count = 0;
1422         struct dma_async_tx_descriptor *tx;
1423         struct async_submit_ctl submit;
1424         int i;
1425
1426         BUG_ON(sh->batch_head);
1427
1428         pr_debug("%s: stripe %llu block: %d\n",
1429                 __func__, (unsigned long long)sh->sector, target);
1430         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1431
1432         for (i = disks; i--; )
1433                 if (i != target)
1434                         xor_srcs[count++] = sh->dev[i].page;
1435
1436         atomic_inc(&sh->count);
1437
1438         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1439                           ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1440         if (unlikely(count == 1))
1441                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1442         else
1443                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1444
1445         return tx;
1446 }
1447
1448 /* set_syndrome_sources - populate source buffers for gen_syndrome
1449  * @srcs - (struct page *) array of size sh->disks
1450  * @sh - stripe_head to parse
1451  *
1452  * Populates srcs in proper layout order for the stripe and returns the
1453  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
1454  * destination buffer is recorded in srcs[count] and the Q destination
1455  * is recorded in srcs[count+1]].
1456  */
1457 static int set_syndrome_sources(struct page **srcs,
1458                                 struct stripe_head *sh,
1459                                 int srctype)
1460 {
1461         int disks = sh->disks;
1462         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1463         int d0_idx = raid6_d0(sh);
1464         int count;
1465         int i;
1466
1467         for (i = 0; i < disks; i++)
1468                 srcs[i] = NULL;
1469
1470         count = 0;
1471         i = d0_idx;
1472         do {
1473                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1474                 struct r5dev *dev = &sh->dev[i];
1475
1476                 if (i == sh->qd_idx || i == sh->pd_idx ||
1477                     (srctype == SYNDROME_SRC_ALL) ||
1478                     (srctype == SYNDROME_SRC_WANT_DRAIN &&
1479                      (test_bit(R5_Wantdrain, &dev->flags) ||
1480                       test_bit(R5_InJournal, &dev->flags))) ||
1481                     (srctype == SYNDROME_SRC_WRITTEN &&
1482                      (dev->written ||
1483                       test_bit(R5_InJournal, &dev->flags)))) {
1484                         if (test_bit(R5_InJournal, &dev->flags))
1485                                 srcs[slot] = sh->dev[i].orig_page;
1486                         else
1487                                 srcs[slot] = sh->dev[i].page;
1488                 }
1489                 i = raid6_next_disk(i, disks);
1490         } while (i != d0_idx);
1491
1492         return syndrome_disks;
1493 }
1494
1495 static struct dma_async_tx_descriptor *
1496 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1497 {
1498         int disks = sh->disks;
1499         struct page **blocks = to_addr_page(percpu, 0);
1500         int target;
1501         int qd_idx = sh->qd_idx;
1502         struct dma_async_tx_descriptor *tx;
1503         struct async_submit_ctl submit;
1504         struct r5dev *tgt;
1505         struct page *dest;
1506         int i;
1507         int count;
1508
1509         BUG_ON(sh->batch_head);
1510         if (sh->ops.target < 0)
1511                 target = sh->ops.target2;
1512         else if (sh->ops.target2 < 0)
1513                 target = sh->ops.target;
1514         else
1515                 /* we should only have one valid target */
1516                 BUG();
1517         BUG_ON(target < 0);
1518         pr_debug("%s: stripe %llu block: %d\n",
1519                 __func__, (unsigned long long)sh->sector, target);
1520
1521         tgt = &sh->dev[target];
1522         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1523         dest = tgt->page;
1524
1525         atomic_inc(&sh->count);
1526
1527         if (target == qd_idx) {
1528                 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1529                 blocks[count] = NULL; /* regenerating p is not necessary */
1530                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1531                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1532                                   ops_complete_compute, sh,
1533                                   to_addr_conv(sh, percpu, 0));
1534                 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1535         } else {
1536                 /* Compute any data- or p-drive using XOR */
1537                 count = 0;
1538                 for (i = disks; i-- ; ) {
1539                         if (i == target || i == qd_idx)
1540                                 continue;
1541                         blocks[count++] = sh->dev[i].page;
1542                 }
1543
1544                 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1545                                   NULL, ops_complete_compute, sh,
1546                                   to_addr_conv(sh, percpu, 0));
1547                 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1548         }
1549
1550         return tx;
1551 }
1552
1553 static struct dma_async_tx_descriptor *
1554 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1555 {
1556         int i, count, disks = sh->disks;
1557         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1558         int d0_idx = raid6_d0(sh);
1559         int faila = -1, failb = -1;
1560         int target = sh->ops.target;
1561         int target2 = sh->ops.target2;
1562         struct r5dev *tgt = &sh->dev[target];
1563         struct r5dev *tgt2 = &sh->dev[target2];
1564         struct dma_async_tx_descriptor *tx;
1565         struct page **blocks = to_addr_page(percpu, 0);
1566         struct async_submit_ctl submit;
1567
1568         BUG_ON(sh->batch_head);
1569         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1570                  __func__, (unsigned long long)sh->sector, target, target2);
1571         BUG_ON(target < 0 || target2 < 0);
1572         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1573         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1574
1575         /* we need to open-code set_syndrome_sources to handle the
1576          * slot number conversion for 'faila' and 'failb'
1577          */
1578         for (i = 0; i < disks ; i++)
1579                 blocks[i] = NULL;
1580         count = 0;
1581         i = d0_idx;
1582         do {
1583                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1584
1585                 blocks[slot] = sh->dev[i].page;
1586
1587                 if (i == target)
1588                         faila = slot;
1589                 if (i == target2)
1590                         failb = slot;
1591                 i = raid6_next_disk(i, disks);
1592         } while (i != d0_idx);
1593
1594         BUG_ON(faila == failb);
1595         if (failb < faila)
1596                 swap(faila, failb);
1597         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1598                  __func__, (unsigned long long)sh->sector, faila, failb);
1599
1600         atomic_inc(&sh->count);
1601
1602         if (failb == syndrome_disks+1) {
1603                 /* Q disk is one of the missing disks */
1604                 if (faila == syndrome_disks) {
1605                         /* Missing P+Q, just recompute */
1606                         init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1607                                           ops_complete_compute, sh,
1608                                           to_addr_conv(sh, percpu, 0));
1609                         return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1610                                                   STRIPE_SIZE, &submit);
1611                 } else {
1612                         struct page *dest;
1613                         int data_target;
1614                         int qd_idx = sh->qd_idx;
1615
1616                         /* Missing D+Q: recompute D from P, then recompute Q */
1617                         if (target == qd_idx)
1618                                 data_target = target2;
1619                         else
1620                                 data_target = target;
1621
1622                         count = 0;
1623                         for (i = disks; i-- ; ) {
1624                                 if (i == data_target || i == qd_idx)
1625                                         continue;
1626                                 blocks[count++] = sh->dev[i].page;
1627                         }
1628                         dest = sh->dev[data_target].page;
1629                         init_async_submit(&submit,
1630                                           ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1631                                           NULL, NULL, NULL,
1632                                           to_addr_conv(sh, percpu, 0));
1633                         tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1634                                        &submit);
1635
1636                         count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1637                         init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1638                                           ops_complete_compute, sh,
1639                                           to_addr_conv(sh, percpu, 0));
1640                         return async_gen_syndrome(blocks, 0, count+2,
1641                                                   STRIPE_SIZE, &submit);
1642                 }
1643         } else {
1644                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1645                                   ops_complete_compute, sh,
1646                                   to_addr_conv(sh, percpu, 0));
1647                 if (failb == syndrome_disks) {
1648                         /* We're missing D+P. */
1649                         return async_raid6_datap_recov(syndrome_disks+2,
1650                                                        STRIPE_SIZE, faila,
1651                                                        blocks, &submit);
1652                 } else {
1653                         /* We're missing D+D. */
1654                         return async_raid6_2data_recov(syndrome_disks+2,
1655                                                        STRIPE_SIZE, faila, failb,
1656                                                        blocks, &submit);
1657                 }
1658         }
1659 }
1660
1661 static void ops_complete_prexor(void *stripe_head_ref)
1662 {
1663         struct stripe_head *sh = stripe_head_ref;
1664
1665         pr_debug("%s: stripe %llu\n", __func__,
1666                 (unsigned long long)sh->sector);
1667
1668         if (r5c_is_writeback(sh->raid_conf->log))
1669                 /*
1670                  * raid5-cache write back uses orig_page during prexor.
1671                  * After prexor, it is time to free orig_page
1672                  */
1673                 r5c_release_extra_page(sh);
1674 }
1675
1676 static struct dma_async_tx_descriptor *
1677 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1678                 struct dma_async_tx_descriptor *tx)
1679 {
1680         int disks = sh->disks;
1681         struct page **xor_srcs = to_addr_page(percpu, 0);
1682         int count = 0, pd_idx = sh->pd_idx, i;
1683         struct async_submit_ctl submit;
1684
1685         /* existing parity data subtracted */
1686         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1687
1688         BUG_ON(sh->batch_head);
1689         pr_debug("%s: stripe %llu\n", __func__,
1690                 (unsigned long long)sh->sector);
1691
1692         for (i = disks; i--; ) {
1693                 struct r5dev *dev = &sh->dev[i];
1694                 /* Only process blocks that are known to be uptodate */
1695                 if (test_bit(R5_InJournal, &dev->flags))
1696                         xor_srcs[count++] = dev->orig_page;
1697                 else if (test_bit(R5_Wantdrain, &dev->flags))
1698                         xor_srcs[count++] = dev->page;
1699         }
1700
1701         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1702                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1703         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1704
1705         return tx;
1706 }
1707
1708 static struct dma_async_tx_descriptor *
1709 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1710                 struct dma_async_tx_descriptor *tx)
1711 {
1712         struct page **blocks = to_addr_page(percpu, 0);
1713         int count;
1714         struct async_submit_ctl submit;
1715
1716         pr_debug("%s: stripe %llu\n", __func__,
1717                 (unsigned long long)sh->sector);
1718
1719         count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1720
1721         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1722                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1723         tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1724
1725         return tx;
1726 }
1727
1728 static struct dma_async_tx_descriptor *
1729 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1730 {
1731         struct r5conf *conf = sh->raid_conf;
1732         int disks = sh->disks;
1733         int i;
1734         struct stripe_head *head_sh = sh;
1735
1736         pr_debug("%s: stripe %llu\n", __func__,
1737                 (unsigned long long)sh->sector);
1738
1739         for (i = disks; i--; ) {
1740                 struct r5dev *dev;
1741                 struct bio *chosen;
1742
1743                 sh = head_sh;
1744                 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1745                         struct bio *wbi;
1746
1747 again:
1748                         dev = &sh->dev[i];
1749                         /*
1750                          * clear R5_InJournal, so when rewriting a page in
1751                          * journal, it is not skipped by r5l_log_stripe()
1752                          */
1753                         clear_bit(R5_InJournal, &dev->flags);
1754                         spin_lock_irq(&sh->stripe_lock);
1755                         chosen = dev->towrite;
1756                         dev->towrite = NULL;
1757                         sh->overwrite_disks = 0;
1758                         BUG_ON(dev->written);
1759                         wbi = dev->written = chosen;
1760                         spin_unlock_irq(&sh->stripe_lock);
1761                         WARN_ON(dev->page != dev->orig_page);
1762
1763                         while (wbi && wbi->bi_iter.bi_sector <
1764                                 dev->sector + STRIPE_SECTORS) {
1765                                 if (wbi->bi_opf & REQ_FUA)
1766                                         set_bit(R5_WantFUA, &dev->flags);
1767                                 if (wbi->bi_opf & REQ_SYNC)
1768                                         set_bit(R5_SyncIO, &dev->flags);
1769                                 if (bio_op(wbi) == REQ_OP_DISCARD)
1770                                         set_bit(R5_Discard, &dev->flags);
1771                                 else {
1772                                         tx = async_copy_data(1, wbi, &dev->page,
1773                                                              dev->sector, tx, sh,
1774                                                              r5c_is_writeback(conf->log));
1775                                         if (dev->page != dev->orig_page &&
1776                                             !r5c_is_writeback(conf->log)) {
1777                                                 set_bit(R5_SkipCopy, &dev->flags);
1778                                                 clear_bit(R5_UPTODATE, &dev->flags);
1779                                                 clear_bit(R5_OVERWRITE, &dev->flags);
1780                                         }
1781                                 }
1782                                 wbi = r5_next_bio(wbi, dev->sector);
1783                         }
1784
1785                         if (head_sh->batch_head) {
1786                                 sh = list_first_entry(&sh->batch_list,
1787                                                       struct stripe_head,
1788                                                       batch_list);
1789                                 if (sh == head_sh)
1790                                         continue;
1791                                 goto again;
1792                         }
1793                 }
1794         }
1795
1796         return tx;
1797 }
1798
1799 static void ops_complete_reconstruct(void *stripe_head_ref)
1800 {
1801         struct stripe_head *sh = stripe_head_ref;
1802         int disks = sh->disks;
1803         int pd_idx = sh->pd_idx;
1804         int qd_idx = sh->qd_idx;
1805         int i;
1806         bool fua = false, sync = false, discard = false;
1807
1808         pr_debug("%s: stripe %llu\n", __func__,
1809                 (unsigned long long)sh->sector);
1810
1811         for (i = disks; i--; ) {
1812                 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1813                 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1814                 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1815         }
1816
1817         for (i = disks; i--; ) {
1818                 struct r5dev *dev = &sh->dev[i];
1819
1820                 if (dev->written || i == pd_idx || i == qd_idx) {
1821                         if (!discard && !test_bit(R5_SkipCopy, &dev->flags)) {
1822                                 set_bit(R5_UPTODATE, &dev->flags);
1823                                 if (test_bit(STRIPE_EXPAND_READY, &sh->state))
1824                                         set_bit(R5_Expanded, &dev->flags);
1825                         }
1826                         if (fua)
1827                                 set_bit(R5_WantFUA, &dev->flags);
1828                         if (sync)
1829                                 set_bit(R5_SyncIO, &dev->flags);
1830                 }
1831         }
1832
1833         if (sh->reconstruct_state == reconstruct_state_drain_run)
1834                 sh->reconstruct_state = reconstruct_state_drain_result;
1835         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1836                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1837         else {
1838                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1839                 sh->reconstruct_state = reconstruct_state_result;
1840         }
1841
1842         set_bit(STRIPE_HANDLE, &sh->state);
1843         raid5_release_stripe(sh);
1844 }
1845
1846 static void
1847 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1848                      struct dma_async_tx_descriptor *tx)
1849 {
1850         int disks = sh->disks;
1851         struct page **xor_srcs;
1852         struct async_submit_ctl submit;
1853         int count, pd_idx = sh->pd_idx, i;
1854         struct page *xor_dest;
1855         int prexor = 0;
1856         unsigned long flags;
1857         int j = 0;
1858         struct stripe_head *head_sh = sh;
1859         int last_stripe;
1860
1861         pr_debug("%s: stripe %llu\n", __func__,
1862                 (unsigned long long)sh->sector);
1863
1864         for (i = 0; i < sh->disks; i++) {
1865                 if (pd_idx == i)
1866                         continue;
1867                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1868                         break;
1869         }
1870         if (i >= sh->disks) {
1871                 atomic_inc(&sh->count);
1872                 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1873                 ops_complete_reconstruct(sh);
1874                 return;
1875         }
1876 again:
1877         count = 0;
1878         xor_srcs = to_addr_page(percpu, j);
1879         /* check if prexor is active which means only process blocks
1880          * that are part of a read-modify-write (written)
1881          */
1882         if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1883                 prexor = 1;
1884                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1885                 for (i = disks; i--; ) {
1886                         struct r5dev *dev = &sh->dev[i];
1887                         if (head_sh->dev[i].written ||
1888                             test_bit(R5_InJournal, &head_sh->dev[i].flags))
1889                                 xor_srcs[count++] = dev->page;
1890                 }
1891         } else {
1892                 xor_dest = sh->dev[pd_idx].page;
1893                 for (i = disks; i--; ) {
1894                         struct r5dev *dev = &sh->dev[i];
1895                         if (i != pd_idx)
1896                                 xor_srcs[count++] = dev->page;
1897                 }
1898         }
1899
1900         /* 1/ if we prexor'd then the dest is reused as a source
1901          * 2/ if we did not prexor then we are redoing the parity
1902          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1903          * for the synchronous xor case
1904          */
1905         last_stripe = !head_sh->batch_head ||
1906                 list_first_entry(&sh->batch_list,
1907                                  struct stripe_head, batch_list) == head_sh;
1908         if (last_stripe) {
1909                 flags = ASYNC_TX_ACK |
1910                         (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1911
1912                 atomic_inc(&head_sh->count);
1913                 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1914                                   to_addr_conv(sh, percpu, j));
1915         } else {
1916                 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1917                 init_async_submit(&submit, flags, tx, NULL, NULL,
1918                                   to_addr_conv(sh, percpu, j));
1919         }
1920
1921         if (unlikely(count == 1))
1922                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1923         else
1924                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1925         if (!last_stripe) {
1926                 j++;
1927                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1928                                       batch_list);
1929                 goto again;
1930         }
1931 }
1932
1933 static void
1934 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1935                      struct dma_async_tx_descriptor *tx)
1936 {
1937         struct async_submit_ctl submit;
1938         struct page **blocks;
1939         int count, i, j = 0;
1940         struct stripe_head *head_sh = sh;
1941         int last_stripe;
1942         int synflags;
1943         unsigned long txflags;
1944
1945         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1946
1947         for (i = 0; i < sh->disks; i++) {
1948                 if (sh->pd_idx == i || sh->qd_idx == i)
1949                         continue;
1950                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1951                         break;
1952         }
1953         if (i >= sh->disks) {
1954                 atomic_inc(&sh->count);
1955                 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1956                 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1957                 ops_complete_reconstruct(sh);
1958                 return;
1959         }
1960
1961 again:
1962         blocks = to_addr_page(percpu, j);
1963
1964         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1965                 synflags = SYNDROME_SRC_WRITTEN;
1966                 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1967         } else {
1968                 synflags = SYNDROME_SRC_ALL;
1969                 txflags = ASYNC_TX_ACK;
1970         }
1971
1972         count = set_syndrome_sources(blocks, sh, synflags);
1973         last_stripe = !head_sh->batch_head ||
1974                 list_first_entry(&sh->batch_list,
1975                                  struct stripe_head, batch_list) == head_sh;
1976
1977         if (last_stripe) {
1978                 atomic_inc(&head_sh->count);
1979                 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1980                                   head_sh, to_addr_conv(sh, percpu, j));
1981         } else
1982                 init_async_submit(&submit, 0, tx, NULL, NULL,
1983                                   to_addr_conv(sh, percpu, j));
1984         tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1985         if (!last_stripe) {
1986                 j++;
1987                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1988                                       batch_list);
1989                 goto again;
1990         }
1991 }
1992
1993 static void ops_complete_check(void *stripe_head_ref)
1994 {
1995         struct stripe_head *sh = stripe_head_ref;
1996
1997         pr_debug("%s: stripe %llu\n", __func__,
1998                 (unsigned long long)sh->sector);
1999
2000         sh->check_state = check_state_check_result;
2001         set_bit(STRIPE_HANDLE, &sh->state);
2002         raid5_release_stripe(sh);
2003 }
2004
2005 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
2006 {
2007         int disks = sh->disks;
2008         int pd_idx = sh->pd_idx;
2009         int qd_idx = sh->qd_idx;
2010         struct page *xor_dest;
2011         struct page **xor_srcs = to_addr_page(percpu, 0);
2012         struct dma_async_tx_descriptor *tx;
2013         struct async_submit_ctl submit;
2014         int count;
2015         int i;
2016
2017         pr_debug("%s: stripe %llu\n", __func__,
2018                 (unsigned long long)sh->sector);
2019
2020         BUG_ON(sh->batch_head);
2021         count = 0;
2022         xor_dest = sh->dev[pd_idx].page;
2023         xor_srcs[count++] = xor_dest;
2024         for (i = disks; i--; ) {
2025                 if (i == pd_idx || i == qd_idx)
2026                         continue;
2027                 xor_srcs[count++] = sh->dev[i].page;
2028         }
2029
2030         init_async_submit(&submit, 0, NULL, NULL, NULL,
2031                           to_addr_conv(sh, percpu, 0));
2032         tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
2033                            &sh->ops.zero_sum_result, &submit);
2034
2035         atomic_inc(&sh->count);
2036         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
2037         tx = async_trigger_callback(&submit);
2038 }
2039
2040 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
2041 {
2042         struct page **srcs = to_addr_page(percpu, 0);
2043         struct async_submit_ctl submit;
2044         int count;
2045
2046         pr_debug("%s: stripe %llu checkp: %d\n", __func__,
2047                 (unsigned long long)sh->sector, checkp);
2048
2049         BUG_ON(sh->batch_head);
2050         count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
2051         if (!checkp)
2052                 srcs[count] = NULL;
2053
2054         atomic_inc(&sh->count);
2055         init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
2056                           sh, to_addr_conv(sh, percpu, 0));
2057         async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
2058                            &sh->ops.zero_sum_result, percpu->spare_page, &submit);
2059 }
2060
2061 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
2062 {
2063         int overlap_clear = 0, i, disks = sh->disks;
2064         struct dma_async_tx_descriptor *tx = NULL;
2065         struct r5conf *conf = sh->raid_conf;
2066         int level = conf->level;
2067         struct raid5_percpu *percpu;
2068         unsigned long cpu;
2069
2070         cpu = get_cpu();
2071         percpu = per_cpu_ptr(conf->percpu, cpu);
2072         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
2073                 ops_run_biofill(sh);
2074                 overlap_clear++;
2075         }
2076
2077         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
2078                 if (level < 6)
2079                         tx = ops_run_compute5(sh, percpu);
2080                 else {
2081                         if (sh->ops.target2 < 0 || sh->ops.target < 0)
2082                                 tx = ops_run_compute6_1(sh, percpu);
2083                         else
2084                                 tx = ops_run_compute6_2(sh, percpu);
2085                 }
2086                 /* terminate the chain if reconstruct is not set to be run */
2087                 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
2088                         async_tx_ack(tx);
2089         }
2090
2091         if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
2092                 if (level < 6)
2093                         tx = ops_run_prexor5(sh, percpu, tx);
2094                 else
2095                         tx = ops_run_prexor6(sh, percpu, tx);
2096         }
2097
2098         if (test_bit(STRIPE_OP_PARTIAL_PARITY, &ops_request))
2099                 tx = ops_run_partial_parity(sh, percpu, tx);
2100
2101         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
2102                 tx = ops_run_biodrain(sh, tx);
2103                 overlap_clear++;
2104         }
2105
2106         if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
2107                 if (level < 6)
2108                         ops_run_reconstruct5(sh, percpu, tx);
2109                 else
2110                         ops_run_reconstruct6(sh, percpu, tx);
2111         }
2112
2113         if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
2114                 if (sh->check_state == check_state_run)
2115                         ops_run_check_p(sh, percpu);
2116                 else if (sh->check_state == check_state_run_q)
2117                         ops_run_check_pq(sh, percpu, 0);
2118                 else if (sh->check_state == check_state_run_pq)
2119                         ops_run_check_pq(sh, percpu, 1);
2120                 else
2121                         BUG();
2122         }
2123
2124         if (overlap_clear && !sh->batch_head)
2125                 for (i = disks; i--; ) {
2126                         struct r5dev *dev = &sh->dev[i];
2127                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
2128                                 wake_up(&sh->raid_conf->wait_for_overlap);
2129                 }
2130         put_cpu();
2131 }
2132
2133 static void free_stripe(struct kmem_cache *sc, struct stripe_head *sh)
2134 {
2135         if (sh->ppl_page)
2136                 __free_page(sh->ppl_page);
2137         kmem_cache_free(sc, sh);
2138 }
2139
2140 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
2141         int disks, struct r5conf *conf)
2142 {
2143         struct stripe_head *sh;
2144         int i;
2145
2146         sh = kmem_cache_zalloc(sc, gfp);
2147         if (sh) {
2148                 spin_lock_init(&sh->stripe_lock);
2149                 spin_lock_init(&sh->batch_lock);
2150                 INIT_LIST_HEAD(&sh->batch_list);
2151                 INIT_LIST_HEAD(&sh->lru);
2152                 INIT_LIST_HEAD(&sh->r5c);
2153                 INIT_LIST_HEAD(&sh->log_list);
2154                 atomic_set(&sh->count, 1);
2155                 sh->raid_conf = conf;
2156                 sh->log_start = MaxSector;
2157                 for (i = 0; i < disks; i++) {
2158                         struct r5dev *dev = &sh->dev[i];
2159
2160                         bio_init(&dev->req, &dev->vec, 1);
2161                         bio_init(&dev->rreq, &dev->rvec, 1);
2162                 }
2163
2164                 if (raid5_has_ppl(conf)) {
2165                         sh->ppl_page = alloc_page(gfp);
2166                         if (!sh->ppl_page) {
2167                                 free_stripe(sc, sh);
2168                                 sh = NULL;
2169                         }
2170                 }
2171         }
2172         return sh;
2173 }
2174 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2175 {
2176         struct stripe_head *sh;
2177
2178         sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size, conf);
2179         if (!sh)
2180                 return 0;
2181
2182         if (grow_buffers(sh, gfp)) {
2183                 shrink_buffers(sh);
2184                 free_stripe(conf->slab_cache, sh);
2185                 return 0;
2186         }
2187         sh->hash_lock_index =
2188                 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2189         /* we just created an active stripe so... */
2190         atomic_inc(&conf->active_stripes);
2191
2192         raid5_release_stripe(sh);
2193         conf->max_nr_stripes++;
2194         return 1;
2195 }
2196
2197 static int grow_stripes(struct r5conf *conf, int num)
2198 {
2199         struct kmem_cache *sc;
2200         int devs = max(conf->raid_disks, conf->previous_raid_disks);
2201
2202         if (conf->mddev->gendisk)
2203                 sprintf(conf->cache_name[0],
2204                         "raid%d-%s", conf->level, mdname(conf->mddev));
2205         else
2206                 sprintf(conf->cache_name[0],
2207                         "raid%d-%p", conf->level, conf->mddev);
2208         sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
2209
2210         conf->active_name = 0;
2211         sc = kmem_cache_create(conf->cache_name[conf->active_name],
2212                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2213                                0, 0, NULL);
2214         if (!sc)
2215                 return 1;
2216         conf->slab_cache = sc;
2217         conf->pool_size = devs;
2218         while (num--)
2219                 if (!grow_one_stripe(conf, GFP_KERNEL))
2220                         return 1;
2221
2222         return 0;
2223 }
2224
2225 /**
2226  * scribble_len - return the required size of the scribble region
2227  * @num - total number of disks in the array
2228  *
2229  * The size must be enough to contain:
2230  * 1/ a struct page pointer for each device in the array +2
2231  * 2/ room to convert each entry in (1) to its corresponding dma
2232  *    (dma_map_page()) or page (page_address()) address.
2233  *
2234  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2235  * calculate over all devices (not just the data blocks), using zeros in place
2236  * of the P and Q blocks.
2237  */
2238 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2239 {
2240         struct flex_array *ret;
2241         size_t len;
2242
2243         len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2244         ret = flex_array_alloc(len, cnt, flags);
2245         if (!ret)
2246                 return NULL;
2247         /* always prealloc all elements, so no locking is required */
2248         if (flex_array_prealloc(ret, 0, cnt, flags)) {
2249                 flex_array_free(ret);
2250                 return NULL;
2251         }
2252         return ret;
2253 }
2254
2255 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2256 {
2257         unsigned long cpu;
2258         int err = 0;
2259
2260         /*
2261          * Never shrink. And mddev_suspend() could deadlock if this is called
2262          * from raid5d. In that case, scribble_disks and scribble_sectors
2263          * should equal to new_disks and new_sectors
2264          */
2265         if (conf->scribble_disks >= new_disks &&
2266             conf->scribble_sectors >= new_sectors)
2267                 return 0;
2268         mddev_suspend(conf->mddev);
2269         get_online_cpus();
2270         for_each_present_cpu(cpu) {
2271                 struct raid5_percpu *percpu;
2272                 struct flex_array *scribble;
2273
2274                 percpu = per_cpu_ptr(conf->percpu, cpu);
2275                 scribble = scribble_alloc(new_disks,
2276                                           new_sectors / STRIPE_SECTORS,
2277                                           GFP_NOIO);
2278
2279                 if (scribble) {
2280                         flex_array_free(percpu->scribble);
2281                         percpu->scribble = scribble;
2282                 } else {
2283                         err = -ENOMEM;
2284                         break;
2285                 }
2286         }
2287         put_online_cpus();
2288         mddev_resume(conf->mddev);
2289         if (!err) {
2290                 conf->scribble_disks = new_disks;
2291                 conf->scribble_sectors = new_sectors;
2292         }
2293         return err;
2294 }
2295
2296 static int resize_stripes(struct r5conf *conf, int newsize)
2297 {
2298         /* Make all the stripes able to hold 'newsize' devices.
2299          * New slots in each stripe get 'page' set to a new page.
2300          *
2301          * This happens in stages:
2302          * 1/ create a new kmem_cache and allocate the required number of
2303          *    stripe_heads.
2304          * 2/ gather all the old stripe_heads and transfer the pages across
2305          *    to the new stripe_heads.  This will have the side effect of
2306          *    freezing the array as once all stripe_heads have been collected,
2307          *    no IO will be possible.  Old stripe heads are freed once their
2308          *    pages have been transferred over, and the old kmem_cache is
2309          *    freed when all stripes are done.
2310          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
2311          *    we simple return a failure status - no need to clean anything up.
2312          * 4/ allocate new pages for the new slots in the new stripe_heads.
2313          *    If this fails, we don't bother trying the shrink the
2314          *    stripe_heads down again, we just leave them as they are.
2315          *    As each stripe_head is processed the new one is released into
2316          *    active service.
2317          *
2318          * Once step2 is started, we cannot afford to wait for a write,
2319          * so we use GFP_NOIO allocations.
2320          */
2321         struct stripe_head *osh, *nsh;
2322         LIST_HEAD(newstripes);
2323         struct disk_info *ndisks;
2324         int err = 0;
2325         struct kmem_cache *sc;
2326         int i;
2327         int hash, cnt;
2328
2329         md_allow_write(conf->mddev);
2330
2331         /* Step 1 */
2332         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2333                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2334                                0, 0, NULL);
2335         if (!sc)
2336                 return -ENOMEM;
2337
2338         /* Need to ensure auto-resizing doesn't interfere */
2339         mutex_lock(&conf->cache_size_mutex);
2340
2341         for (i = conf->max_nr_stripes; i; i--) {
2342                 nsh = alloc_stripe(sc, GFP_KERNEL, newsize, conf);
2343                 if (!nsh)
2344                         break;
2345
2346                 list_add(&nsh->lru, &newstripes);
2347         }
2348         if (i) {
2349                 /* didn't get enough, give up */
2350                 while (!list_empty(&newstripes)) {
2351                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
2352                         list_del(&nsh->lru);
2353                         free_stripe(sc, nsh);
2354                 }
2355                 kmem_cache_destroy(sc);
2356                 mutex_unlock(&conf->cache_size_mutex);
2357                 return -ENOMEM;
2358         }
2359         /* Step 2 - Must use GFP_NOIO now.
2360          * OK, we have enough stripes, start collecting inactive
2361          * stripes and copying them over
2362          */
2363         hash = 0;
2364         cnt = 0;
2365         list_for_each_entry(nsh, &newstripes, lru) {
2366                 lock_device_hash_lock(conf, hash);
2367                 wait_event_cmd(conf->wait_for_stripe,
2368                                     !list_empty(conf->inactive_list + hash),
2369                                     unlock_device_hash_lock(conf, hash),
2370                                     lock_device_hash_lock(conf, hash));
2371                 osh = get_free_stripe(conf, hash);
2372                 unlock_device_hash_lock(conf, hash);
2373
2374                 for(i=0; i<conf->pool_size; i++) {
2375                         nsh->dev[i].page = osh->dev[i].page;
2376                         nsh->dev[i].orig_page = osh->dev[i].page;
2377                 }
2378                 nsh->hash_lock_index = hash;
2379                 free_stripe(conf->slab_cache, osh);
2380                 cnt++;
2381                 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2382                     !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2383                         hash++;
2384                         cnt = 0;
2385                 }
2386         }
2387         kmem_cache_destroy(conf->slab_cache);
2388
2389         /* Step 3.
2390          * At this point, we are holding all the stripes so the array
2391          * is completely stalled, so now is a good time to resize
2392          * conf->disks and the scribble region
2393          */
2394         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2395         if (ndisks) {
2396                 for (i = 0; i < conf->pool_size; i++)
2397                         ndisks[i] = conf->disks[i];
2398
2399                 for (i = conf->pool_size; i < newsize; i++) {
2400                         ndisks[i].extra_page = alloc_page(GFP_NOIO);
2401                         if (!ndisks[i].extra_page)
2402                                 err = -ENOMEM;
2403                 }
2404
2405                 if (err) {
2406                         for (i = conf->pool_size; i < newsize; i++)
2407                                 if (ndisks[i].extra_page)
2408                                         put_page(ndisks[i].extra_page);
2409                         kfree(ndisks);
2410                 } else {
2411                         kfree(conf->disks);
2412                         conf->disks = ndisks;
2413                 }
2414         } else
2415                 err = -ENOMEM;
2416
2417         mutex_unlock(&conf->cache_size_mutex);
2418
2419         conf->slab_cache = sc;
2420         conf->active_name = 1-conf->active_name;
2421
2422         /* Step 4, return new stripes to service */
2423         while(!list_empty(&newstripes)) {
2424                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2425                 list_del_init(&nsh->lru);
2426
2427                 for (i=conf->raid_disks; i < newsize; i++)
2428                         if (nsh->dev[i].page == NULL) {
2429                                 struct page *p = alloc_page(GFP_NOIO);
2430                                 nsh->dev[i].page = p;
2431                                 nsh->dev[i].orig_page = p;
2432                                 if (!p)
2433                                         err = -ENOMEM;
2434                         }
2435                 raid5_release_stripe(nsh);
2436         }
2437         /* critical section pass, GFP_NOIO no longer needed */
2438
2439         if (!err)
2440                 conf->pool_size = newsize;
2441         return err;
2442 }
2443
2444 static int drop_one_stripe(struct r5conf *conf)
2445 {
2446         struct stripe_head *sh;
2447         int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2448
2449         spin_lock_irq(conf->hash_locks + hash);
2450         sh = get_free_stripe(conf, hash);
2451         spin_unlock_irq(conf->hash_locks + hash);
2452         if (!sh)
2453                 return 0;
2454         BUG_ON(atomic_read(&sh->count));
2455         shrink_buffers(sh);
2456         free_stripe(conf->slab_cache, sh);
2457         atomic_dec(&conf->active_stripes);
2458         conf->max_nr_stripes--;
2459         return 1;
2460 }
2461
2462 static void shrink_stripes(struct r5conf *conf)
2463 {
2464         while (conf->max_nr_stripes &&
2465                drop_one_stripe(conf))
2466                 ;
2467
2468         kmem_cache_destroy(conf->slab_cache);
2469         conf->slab_cache = NULL;
2470 }
2471
2472 static void raid5_end_read_request(struct bio * bi)
2473 {
2474         struct stripe_head *sh = bi->bi_private;
2475         struct r5conf *conf = sh->raid_conf;
2476         int disks = sh->disks, i;
2477         char b[BDEVNAME_SIZE];
2478         struct md_rdev *rdev = NULL;
2479         sector_t s;
2480
2481         for (i=0 ; i<disks; i++)
2482                 if (bi == &sh->dev[i].req)
2483                         break;
2484
2485         pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2486                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2487                 bi->bi_status);
2488         if (i == disks) {
2489                 bio_reset(bi);
2490                 BUG();
2491                 return;
2492         }
2493         if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2494                 /* If replacement finished while this request was outstanding,
2495                  * 'replacement' might be NULL already.
2496                  * In that case it moved down to 'rdev'.
2497                  * rdev is not removed until all requests are finished.
2498                  */
2499                 rdev = conf->disks[i].replacement;
2500         if (!rdev)
2501                 rdev = conf->disks[i].rdev;
2502
2503         if (use_new_offset(conf, sh))
2504                 s = sh->sector + rdev->new_data_offset;
2505         else
2506                 s = sh->sector + rdev->data_offset;
2507         if (!bi->bi_status) {
2508                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2509                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2510                         /* Note that this cannot happen on a
2511                          * replacement device.  We just fail those on
2512                          * any error
2513                          */
2514                         pr_info_ratelimited(
2515                                 "md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2516                                 mdname(conf->mddev), STRIPE_SECTORS,
2517                                 (unsigned long long)s,
2518                                 bdevname(rdev->bdev, b));
2519                         atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2520                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2521                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2522                 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2523                         clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2524
2525                 if (test_bit(R5_InJournal, &sh->dev[i].flags))
2526                         /*
2527                          * end read for a page in journal, this
2528                          * must be preparing for prexor in rmw
2529                          */
2530                         set_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2531
2532                 if (atomic_read(&rdev->read_errors))
2533                         atomic_set(&rdev->read_errors, 0);
2534         } else {
2535                 const char *bdn = bdevname(rdev->bdev, b);
2536                 int retry = 0;
2537                 int set_bad = 0;
2538
2539                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2540                 atomic_inc(&rdev->read_errors);
2541                 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2542                         pr_warn_ratelimited(
2543                                 "md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2544                                 mdname(conf->mddev),
2545                                 (unsigned long long)s,
2546                                 bdn);
2547                 else if (conf->mddev->degraded >= conf->max_degraded) {
2548                         set_bad = 1;
2549                         pr_warn_ratelimited(
2550                                 "md/raid:%s: read error not correctable (sector %llu on %s).\n",
2551                                 mdname(conf->mddev),
2552                                 (unsigned long long)s,
2553                                 bdn);
2554                 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2555                         /* Oh, no!!! */
2556                         set_bad = 1;
2557                         pr_warn_ratelimited(
2558                                 "md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2559                                 mdname(conf->mddev),
2560                                 (unsigned long long)s,
2561                                 bdn);
2562                 } else if (atomic_read(&rdev->read_errors)
2563                          > conf->max_nr_stripes)
2564                         pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2565                                mdname(conf->mddev), bdn);
2566                 else
2567                         retry = 1;
2568                 if (set_bad && test_bit(In_sync, &rdev->flags)
2569                     && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2570                         retry = 1;
2571                 if (retry)
2572                         if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2573                                 set_bit(R5_ReadError, &sh->dev[i].flags);
2574                                 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2575                         } else
2576                                 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2577                 else {
2578                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2579                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2580                         if (!(set_bad
2581                               && test_bit(In_sync, &rdev->flags)
2582                               && rdev_set_badblocks(
2583                                       rdev, sh->sector, STRIPE_SECTORS, 0)))
2584                                 md_error(conf->mddev, rdev);
2585                 }
2586         }
2587         rdev_dec_pending(rdev, conf->mddev);
2588         bio_reset(bi);
2589         clear_bit(R5_LOCKED, &sh->dev[i].flags);
2590         set_bit(STRIPE_HANDLE, &sh->state);
2591         raid5_release_stripe(sh);
2592 }
2593
2594 static void raid5_end_write_request(struct bio *bi)
2595 {
2596         struct stripe_head *sh = bi->bi_private;
2597         struct r5conf *conf = sh->raid_conf;
2598         int disks = sh->disks, i;
2599         struct md_rdev *uninitialized_var(rdev);
2600         sector_t first_bad;
2601         int bad_sectors;
2602         int replacement = 0;
2603
2604         for (i = 0 ; i < disks; i++) {
2605                 if (bi == &sh->dev[i].req) {
2606                         rdev = conf->disks[i].rdev;
2607                         break;
2608                 }
2609                 if (bi == &sh->dev[i].rreq) {
2610                         rdev = conf->disks[i].replacement;
2611                         if (rdev)
2612                                 replacement = 1;
2613                         else
2614                                 /* rdev was removed and 'replacement'
2615                                  * replaced it.  rdev is not removed
2616                                  * until all requests are finished.
2617                                  */
2618                                 rdev = conf->disks[i].rdev;
2619                         break;
2620                 }
2621         }
2622         pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2623                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2624                 bi->bi_status);
2625         if (i == disks) {
2626                 bio_reset(bi);
2627                 BUG();
2628                 return;
2629         }
2630
2631         if (replacement) {
2632                 if (bi->bi_status)
2633                         md_error(conf->mddev, rdev);
2634                 else if (is_badblock(rdev, sh->sector,
2635                                      STRIPE_SECTORS,
2636                                      &first_bad, &bad_sectors))
2637                         set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2638         } else {
2639                 if (bi->bi_status) {
2640                         set_bit(STRIPE_DEGRADED, &sh->state);
2641                         set_bit(WriteErrorSeen, &rdev->flags);
2642                         set_bit(R5_WriteError, &sh->dev[i].flags);
2643                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
2644                                 set_bit(MD_RECOVERY_NEEDED,
2645                                         &rdev->mddev->recovery);
2646                 } else if (is_badblock(rdev, sh->sector,
2647                                        STRIPE_SECTORS,
2648                                        &first_bad, &bad_sectors)) {
2649                         set_bit(R5_MadeGood, &sh->dev[i].flags);
2650                         if (test_bit(R5_ReadError, &sh->dev[i].flags))
2651                                 /* That was a successful write so make
2652                                  * sure it looks like we already did
2653                                  * a re-write.
2654                                  */
2655                                 set_bit(R5_ReWrite, &sh->dev[i].flags);
2656                 }
2657         }
2658         rdev_dec_pending(rdev, conf->mddev);
2659
2660         if (sh->batch_head && bi->bi_status && !replacement)
2661                 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2662
2663         bio_reset(bi);
2664         if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2665                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2666         set_bit(STRIPE_HANDLE, &sh->state);
2667         raid5_release_stripe(sh);
2668
2669         if (sh->batch_head && sh != sh->batch_head)
2670                 raid5_release_stripe(sh->batch_head);
2671 }
2672
2673 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2674 {
2675         char b[BDEVNAME_SIZE];
2676         struct r5conf *conf = mddev->private;
2677         unsigned long flags;
2678         pr_debug("raid456: error called\n");
2679
2680         spin_lock_irqsave(&conf->device_lock, flags);
2681         clear_bit(In_sync, &rdev->flags);
2682         mddev->degraded = raid5_calc_degraded(conf);
2683         spin_unlock_irqrestore(&conf->device_lock, flags);
2684         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2685
2686         set_bit(Blocked, &rdev->flags);
2687         set_bit(Faulty, &rdev->flags);
2688         set_mask_bits(&mddev->sb_flags, 0,
2689                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2690         pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
2691                 "md/raid:%s: Operation continuing on %d devices.\n",
2692                 mdname(mddev),
2693                 bdevname(rdev->bdev, b),
2694                 mdname(mddev),
2695                 conf->raid_disks - mddev->degraded);
2696         r5c_update_on_rdev_error(mddev, rdev);
2697 }
2698
2699 /*
2700  * Input: a 'big' sector number,
2701  * Output: index of the data and parity disk, and the sector # in them.
2702  */
2703 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2704                               int previous, int *dd_idx,
2705                               struct stripe_head *sh)
2706 {
2707         sector_t stripe, stripe2;
2708         sector_t chunk_number;
2709         unsigned int chunk_offset;
2710         int pd_idx, qd_idx;
2711         int ddf_layout = 0;
2712         sector_t new_sector;
2713         int algorithm = previous ? conf->prev_algo
2714                                  : conf->algorithm;
2715         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2716                                          : conf->chunk_sectors;
2717         int raid_disks = previous ? conf->previous_raid_disks
2718                                   : conf->raid_disks;
2719         int data_disks = raid_disks - conf->max_degraded;
2720
2721         /* First compute the information on this sector */
2722
2723         /*
2724          * Compute the chunk number and the sector offset inside the chunk
2725          */
2726         chunk_offset = sector_div(r_sector, sectors_per_chunk);
2727         chunk_number = r_sector;
2728
2729         /*
2730          * Compute the stripe number
2731          */
2732         stripe = chunk_number;
2733         *dd_idx = sector_div(stripe, data_disks);
2734         stripe2 = stripe;
2735         /*
2736          * Select the parity disk based on the user selected algorithm.
2737          */
2738         pd_idx = qd_idx = -1;
2739         switch(conf->level) {
2740         case 4:
2741                 pd_idx = data_disks;
2742                 break;
2743         case 5:
2744                 switch (algorithm) {
2745                 case ALGORITHM_LEFT_ASYMMETRIC:
2746                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2747                         if (*dd_idx >= pd_idx)
2748                                 (*dd_idx)++;
2749                         break;
2750                 case ALGORITHM_RIGHT_ASYMMETRIC:
2751                         pd_idx = sector_div(stripe2, raid_disks);
2752                         if (*dd_idx >= pd_idx)
2753                                 (*dd_idx)++;
2754                         break;
2755                 case ALGORITHM_LEFT_SYMMETRIC:
2756                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2757                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2758                         break;
2759                 case ALGORITHM_RIGHT_SYMMETRIC:
2760                         pd_idx = sector_div(stripe2, raid_disks);
2761                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2762                         break;
2763                 case ALGORITHM_PARITY_0:
2764                         pd_idx = 0;
2765                         (*dd_idx)++;
2766                         break;
2767                 case ALGORITHM_PARITY_N:
2768                         pd_idx = data_disks;
2769                         break;
2770                 default:
2771                         BUG();
2772                 }
2773                 break;
2774         case 6:
2775
2776                 switch (algorithm) {
2777                 case ALGORITHM_LEFT_ASYMMETRIC:
2778                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2779                         qd_idx = pd_idx + 1;
2780                         if (pd_idx == raid_disks-1) {
2781                                 (*dd_idx)++;    /* Q D D D P */
2782                                 qd_idx = 0;
2783                         } else if (*dd_idx >= pd_idx)
2784                                 (*dd_idx) += 2; /* D D P Q D */
2785                         break;
2786                 case ALGORITHM_RIGHT_ASYMMETRIC:
2787                         pd_idx = sector_div(stripe2, raid_disks);
2788                         qd_idx = pd_idx + 1;
2789                         if (pd_idx == raid_disks-1) {
2790                                 (*dd_idx)++;    /* Q D D D P */
2791                                 qd_idx = 0;
2792                         } else if (*dd_idx >= pd_idx)
2793                                 (*dd_idx) += 2; /* D D P Q D */
2794                         break;
2795                 case ALGORITHM_LEFT_SYMMETRIC:
2796                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2797                         qd_idx = (pd_idx + 1) % raid_disks;
2798                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2799                         break;
2800                 case ALGORITHM_RIGHT_SYMMETRIC:
2801                         pd_idx = sector_div(stripe2, raid_disks);
2802                         qd_idx = (pd_idx + 1) % raid_disks;
2803                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2804                         break;
2805
2806                 case ALGORITHM_PARITY_0:
2807                         pd_idx = 0;
2808                         qd_idx = 1;
2809                         (*dd_idx) += 2;
2810                         break;
2811                 case ALGORITHM_PARITY_N:
2812                         pd_idx = data_disks;
2813                         qd_idx = data_disks + 1;
2814                         break;
2815
2816                 case ALGORITHM_ROTATING_ZERO_RESTART:
2817                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
2818                          * of blocks for computing Q is different.
2819                          */
2820                         pd_idx = sector_div(stripe2, raid_disks);
2821                         qd_idx = pd_idx + 1;
2822                         if (pd_idx == raid_disks-1) {
2823                                 (*dd_idx)++;    /* Q D D D P */
2824                                 qd_idx = 0;
2825                         } else if (*dd_idx >= pd_idx)
2826                                 (*dd_idx) += 2; /* D D P Q D */
2827                         ddf_layout = 1;
2828                         break;
2829
2830                 case ALGORITHM_ROTATING_N_RESTART:
2831                         /* Same a left_asymmetric, by first stripe is
2832                          * D D D P Q  rather than
2833                          * Q D D D P
2834                          */
2835                         stripe2 += 1;
2836                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2837                         qd_idx = pd_idx + 1;
2838                         if (pd_idx == raid_disks-1) {
2839                                 (*dd_idx)++;    /* Q D D D P */
2840                                 qd_idx = 0;
2841                         } else if (*dd_idx >= pd_idx)
2842                                 (*dd_idx) += 2; /* D D P Q D */
2843                         ddf_layout = 1;
2844                         break;
2845
2846                 case ALGORITHM_ROTATING_N_CONTINUE:
2847                         /* Same as left_symmetric but Q is before P */
2848                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2849                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2850                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2851                         ddf_layout = 1;
2852                         break;
2853
2854                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2855                         /* RAID5 left_asymmetric, with Q on last device */
2856                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2857                         if (*dd_idx >= pd_idx)
2858                                 (*dd_idx)++;
2859                         qd_idx = raid_disks - 1;
2860                         break;
2861
2862                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2863                         pd_idx = sector_div(stripe2, raid_disks-1);
2864                         if (*dd_idx >= pd_idx)
2865                                 (*dd_idx)++;
2866                         qd_idx = raid_disks - 1;
2867                         break;
2868
2869                 case ALGORITHM_LEFT_SYMMETRIC_6:
2870                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2871                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2872                         qd_idx = raid_disks - 1;
2873                         break;
2874
2875                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2876                         pd_idx = sector_div(stripe2, raid_disks-1);
2877                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2878                         qd_idx = raid_disks - 1;
2879                         break;
2880
2881                 case ALGORITHM_PARITY_0_6:
2882                         pd_idx = 0;
2883                         (*dd_idx)++;
2884                         qd_idx = raid_disks - 1;
2885                         break;
2886
2887                 default:
2888                         BUG();
2889                 }
2890                 break;
2891         }
2892
2893         if (sh) {
2894                 sh->pd_idx = pd_idx;
2895                 sh->qd_idx = qd_idx;
2896                 sh->ddf_layout = ddf_layout;
2897         }
2898         /*
2899          * Finally, compute the new sector number
2900          */
2901         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2902         return new_sector;
2903 }
2904
2905 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2906 {
2907         struct r5conf *conf = sh->raid_conf;
2908         int raid_disks = sh->disks;
2909         int data_disks = raid_disks - conf->max_degraded;
2910         sector_t new_sector = sh->sector, check;
2911         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2912                                          : conf->chunk_sectors;
2913         int algorithm = previous ? conf->prev_algo
2914                                  : conf->algorithm;
2915         sector_t stripe;
2916         int chunk_offset;
2917         sector_t chunk_number;
2918         int dummy1, dd_idx = i;
2919         sector_t r_sector;
2920         struct stripe_head sh2;
2921
2922         chunk_offset = sector_div(new_sector, sectors_per_chunk);
2923         stripe = new_sector;
2924
2925         if (i == sh->pd_idx)
2926                 return 0;
2927         switch(conf->level) {
2928         case 4: break;
2929         case 5:
2930                 switch (algorithm) {
2931                 case ALGORITHM_LEFT_ASYMMETRIC:
2932                 case ALGORITHM_RIGHT_ASYMMETRIC:
2933                         if (i > sh->pd_idx)
2934                                 i--;
2935                         break;
2936                 case ALGORITHM_LEFT_SYMMETRIC:
2937                 case ALGORITHM_RIGHT_SYMMETRIC:
2938                         if (i < sh->pd_idx)
2939                                 i += raid_disks;
2940                         i -= (sh->pd_idx + 1);
2941                         break;
2942                 case ALGORITHM_PARITY_0:
2943                         i -= 1;
2944                         break;
2945                 case ALGORITHM_PARITY_N:
2946                         break;
2947                 default:
2948                         BUG();
2949                 }
2950                 break;
2951         case 6:
2952                 if (i == sh->qd_idx)
2953                         return 0; /* It is the Q disk */
2954                 switch (algorithm) {
2955                 case ALGORITHM_LEFT_ASYMMETRIC:
2956                 case ALGORITHM_RIGHT_ASYMMETRIC:
2957                 case ALGORITHM_ROTATING_ZERO_RESTART:
2958                 case ALGORITHM_ROTATING_N_RESTART:
2959                         if (sh->pd_idx == raid_disks-1)
2960                                 i--;    /* Q D D D P */
2961                         else if (i > sh->pd_idx)
2962                                 i -= 2; /* D D P Q D */
2963                         break;
2964                 case ALGORITHM_LEFT_SYMMETRIC:
2965                 case ALGORITHM_RIGHT_SYMMETRIC:
2966                         if (sh->pd_idx == raid_disks-1)
2967                                 i--; /* Q D D D P */
2968                         else {
2969                                 /* D D P Q D */
2970                                 if (i < sh->pd_idx)
2971                                         i += raid_disks;
2972                                 i -= (sh->pd_idx + 2);
2973                         }
2974                         break;
2975                 case ALGORITHM_PARITY_0:
2976                         i -= 2;
2977                         break;
2978                 case ALGORITHM_PARITY_N:
2979                         break;
2980                 case ALGORITHM_ROTATING_N_CONTINUE:
2981                         /* Like left_symmetric, but P is before Q */
2982                         if (sh->pd_idx == 0)
2983                                 i--;    /* P D D D Q */
2984                         else {
2985                                 /* D D Q P D */
2986                                 if (i < sh->pd_idx)
2987                                         i += raid_disks;
2988                                 i -= (sh->pd_idx + 1);
2989                         }
2990                         break;
2991                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2992                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2993                         if (i > sh->pd_idx)
2994                                 i--;
2995                         break;
2996                 case ALGORITHM_LEFT_SYMMETRIC_6:
2997                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2998                         if (i < sh->pd_idx)
2999                                 i += data_disks + 1;
3000                         i -= (sh->pd_idx + 1);
3001                         break;
3002                 case ALGORITHM_PARITY_0_6:
3003                         i -= 1;
3004                         break;
3005                 default:
3006                         BUG();
3007                 }
3008                 break;
3009         }
3010
3011         chunk_number = stripe * data_disks + i;
3012         r_sector = chunk_number * sectors_per_chunk + chunk_offset;
3013
3014         check = raid5_compute_sector(conf, r_sector,
3015                                      previous, &dummy1, &sh2);
3016         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
3017                 || sh2.qd_idx != sh->qd_idx) {
3018                 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3019                         mdname(conf->mddev));
3020                 return 0;
3021         }
3022         return r_sector;
3023 }
3024
3025 /*
3026  * There are cases where we want handle_stripe_dirtying() and
3027  * schedule_reconstruction() to delay towrite to some dev of a stripe.
3028  *
3029  * This function checks whether we want to delay the towrite. Specifically,
3030  * we delay the towrite when:
3031  *
3032  *   1. degraded stripe has a non-overwrite to the missing dev, AND this
3033  *      stripe has data in journal (for other devices).
3034  *
3035  *      In this case, when reading data for the non-overwrite dev, it is
3036  *      necessary to handle complex rmw of write back cache (prexor with
3037  *      orig_page, and xor with page). To keep read path simple, we would
3038  *      like to flush data in journal to RAID disks first, so complex rmw
3039  *      is handled in the write patch (handle_stripe_dirtying).
3040  *
3041  *   2. when journal space is critical (R5C_LOG_CRITICAL=1)
3042  *
3043  *      It is important to be able to flush all stripes in raid5-cache.
3044  *      Therefore, we need reserve some space on the journal device for
3045  *      these flushes. If flush operation includes pending writes to the
3046  *      stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3047  *      for the flush out. If we exclude these pending writes from flush
3048  *      operation, we only need (conf->max_degraded + 1) pages per stripe.
3049  *      Therefore, excluding pending writes in these cases enables more
3050  *      efficient use of the journal device.
3051  *
3052  *      Note: To make sure the stripe makes progress, we only delay
3053  *      towrite for stripes with data already in journal (injournal > 0).
3054  *      When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3055  *      no_space_stripes list.
3056  *
3057  *   3. during journal failure
3058  *      In journal failure, we try to flush all cached data to raid disks
3059  *      based on data in stripe cache. The array is read-only to upper
3060  *      layers, so we would skip all pending writes.
3061  *
3062  */
3063 static inline bool delay_towrite(struct r5conf *conf,
3064                                  struct r5dev *dev,
3065                                  struct stripe_head_state *s)
3066 {
3067         /* case 1 above */
3068         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3069             !test_bit(R5_Insync, &dev->flags) && s->injournal)
3070                 return true;
3071         /* case 2 above */
3072         if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
3073             s->injournal > 0)
3074                 return true;
3075         /* case 3 above */
3076         if (s->log_failed && s->injournal)
3077                 return true;
3078         return false;
3079 }
3080
3081 static void
3082 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
3083                          int rcw, int expand)
3084 {
3085         int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
3086         struct r5conf *conf = sh->raid_conf;
3087         int level = conf->level;
3088
3089         if (rcw) {
3090                 /*
3091                  * In some cases, handle_stripe_dirtying initially decided to
3092                  * run rmw and allocates extra page for prexor. However, rcw is
3093                  * cheaper later on. We need to free the extra page now,
3094                  * because we won't be able to do that in ops_complete_prexor().
3095                  */
3096                 r5c_release_extra_page(sh);
3097
3098                 for (i = disks; i--; ) {
3099                         struct r5dev *dev = &sh->dev[i];
3100
3101                         if (dev->towrite && !delay_towrite(conf, dev, s)) {
3102                                 set_bit(R5_LOCKED, &dev->flags);
3103                                 set_bit(R5_Wantdrain, &dev->flags);
3104                                 if (!expand)
3105                                         clear_bit(R5_UPTODATE, &dev->flags);
3106                                 s->locked++;
3107                         } else if (test_bit(R5_InJournal, &dev->flags)) {
3108                                 set_bit(R5_LOCKED, &dev->flags);
3109                                 s->locked++;
3110                         }
3111                 }
3112                 /* if we are not expanding this is a proper write request, and
3113                  * there will be bios with new data to be drained into the
3114                  * stripe cache
3115                  */
3116                 if (!expand) {
3117                         if (!s->locked)
3118                                 /* False alarm, nothing to do */
3119                                 return;
3120                         sh->reconstruct_state = reconstruct_state_drain_run;
3121                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3122                 } else
3123                         sh->reconstruct_state = reconstruct_state_run;
3124
3125                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3126
3127                 if (s->locked + conf->max_degraded == disks)
3128                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
3129                                 atomic_inc(&conf->pending_full_writes);
3130         } else {
3131                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
3132                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
3133                 BUG_ON(level == 6 &&
3134                         (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
3135                            test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
3136
3137                 for (i = disks; i--; ) {
3138                         struct r5dev *dev = &sh->dev[i];
3139                         if (i == pd_idx || i == qd_idx)
3140                                 continue;
3141
3142                         if (dev->towrite &&
3143                             (test_bit(R5_UPTODATE, &dev->flags) ||
3144                              test_bit(R5_Wantcompute, &dev->flags))) {
3145                                 set_bit(R5_Wantdrain, &dev->flags);
3146                                 set_bit(R5_LOCKED, &dev->flags);
3147                                 clear_bit(R5_UPTODATE, &dev->flags);
3148                                 s->locked++;
3149                         } else if (test_bit(R5_InJournal, &dev->flags)) {
3150                                 set_bit(R5_LOCKED, &dev->flags);
3151                                 s->locked++;
3152                         }
3153                 }
3154                 if (!s->locked)
3155                         /* False alarm - nothing to do */
3156                         return;
3157                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
3158                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
3159                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3160                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3161         }
3162
3163         /* keep the parity disk(s) locked while asynchronous operations
3164          * are in flight
3165          */
3166         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
3167         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3168         s->locked++;
3169
3170         if (level == 6) {
3171                 int qd_idx = sh->qd_idx;
3172                 struct r5dev *dev = &sh->dev[qd_idx];
3173
3174                 set_bit(R5_LOCKED, &dev->flags);
3175                 clear_bit(R5_UPTODATE, &dev->flags);
3176                 s->locked++;
3177         }
3178
3179         if (raid5_has_ppl(sh->raid_conf) && sh->ppl_page &&
3180             test_bit(STRIPE_OP_BIODRAIN, &s->ops_request) &&
3181             !test_bit(STRIPE_FULL_WRITE, &sh->state) &&
3182             test_bit(R5_Insync, &sh->dev[pd_idx].flags))
3183                 set_bit(STRIPE_OP_PARTIAL_PARITY, &s->ops_request);
3184
3185         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3186                 __func__, (unsigned long long)sh->sector,
3187                 s->locked, s->ops_request);
3188 }
3189
3190 /*
3191  * Each stripe/dev can have one or more bion attached.
3192  * toread/towrite point to the first in a chain.
3193  * The bi_next chain must be in order.
3194  */
3195 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
3196                           int forwrite, int previous)
3197 {
3198         struct bio **bip;
3199         struct r5conf *conf = sh->raid_conf;
3200         int firstwrite=0;
3201
3202         pr_debug("adding bi b#%llu to stripe s#%llu\n",
3203                 (unsigned long long)bi->bi_iter.bi_sector,
3204                 (unsigned long long)sh->sector);
3205
3206         spin_lock_irq(&sh->stripe_lock);
3207         /* Don't allow new IO added to stripes in batch list */
3208         if (sh->batch_head)
3209                 goto overlap;
3210         if (forwrite) {
3211                 bip = &sh->dev[dd_idx].towrite;
3212                 if (*bip == NULL)
3213                         firstwrite = 1;
3214         } else
3215                 bip = &sh->dev[dd_idx].toread;
3216         while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3217                 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3218                         goto overlap;
3219                 bip = & (*bip)->bi_next;
3220         }
3221         if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3222                 goto overlap;
3223
3224         if (forwrite && raid5_has_ppl(conf)) {
3225                 /*
3226                  * With PPL only writes to consecutive data chunks within a
3227                  * stripe are allowed because for a single stripe_head we can
3228                  * only have one PPL entry at a time, which describes one data
3229                  * range. Not really an overlap, but wait_for_overlap can be
3230                  * used to handle this.
3231                  */
3232                 sector_t sector;
3233                 sector_t first = 0;
3234                 sector_t last = 0;
3235                 int count = 0;
3236                 int i;
3237
3238                 for (i = 0; i < sh->disks; i++) {
3239                         if (i != sh->pd_idx &&
3240                             (i == dd_idx || sh->dev[i].towrite)) {
3241                                 sector = sh->dev[i].sector;
3242                                 if (count == 0 || sector < first)
3243                                         first = sector;
3244                                 if (sector > last)
3245                                         last = sector;
3246                                 count++;
3247                         }
3248                 }
3249
3250                 if (first + conf->chunk_sectors * (count - 1) != last)
3251                         goto overlap;
3252         }
3253
3254         if (!forwrite || previous)
3255                 clear_bit(STRIPE_BATCH_READY, &sh->state);
3256
3257         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3258         if (*bip)
3259                 bi->bi_next = *bip;
3260         *bip = bi;
3261         bio_inc_remaining(bi);
3262         md_write_inc(conf->mddev, bi);
3263
3264         if (forwrite) {
3265                 /* check if page is covered */
3266                 sector_t sector = sh->dev[dd_idx].sector;
3267                 for (bi=sh->dev[dd_idx].towrite;
3268                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3269                              bi && bi->bi_iter.bi_sector <= sector;
3270                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3271                         if (bio_end_sector(bi) >= sector)
3272                                 sector = bio_end_sector(bi);
3273                 }
3274                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3275                         if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3276                                 sh->overwrite_disks++;
3277         }
3278
3279         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3280                 (unsigned long long)(*bip)->bi_iter.bi_sector,
3281                 (unsigned long long)sh->sector, dd_idx);
3282
3283         if (conf->mddev->bitmap && firstwrite) {
3284                 /* Cannot hold spinlock over bitmap_startwrite,
3285                  * but must ensure this isn't added to a batch until
3286                  * we have added to the bitmap and set bm_seq.
3287                  * So set STRIPE_BITMAP_PENDING to prevent
3288                  * batching.
3289                  * If multiple add_stripe_bio() calls race here they
3290                  * much all set STRIPE_BITMAP_PENDING.  So only the first one
3291                  * to complete "bitmap_startwrite" gets to set
3292                  * STRIPE_BIT_DELAY.  This is important as once a stripe
3293                  * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3294                  * any more.
3295                  */
3296                 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3297                 spin_unlock_irq(&sh->stripe_lock);
3298                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3299                                   STRIPE_SECTORS, 0);
3300                 spin_lock_irq(&sh->stripe_lock);
3301                 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3302                 if (!sh->batch_head) {
3303                         sh->bm_seq = conf->seq_flush+1;
3304                         set_bit(STRIPE_BIT_DELAY, &sh->state);
3305                 }
3306         }
3307         spin_unlock_irq(&sh->stripe_lock);
3308
3309         if (stripe_can_batch(sh))
3310                 stripe_add_to_batch_list(conf, sh);
3311         return 1;
3312
3313  overlap:
3314         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3315         spin_unlock_irq(&sh->stripe_lock);
3316         return 0;
3317 }
3318
3319 static void end_reshape(struct r5conf *conf);
3320
3321 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3322                             struct stripe_head *sh)
3323 {
3324         int sectors_per_chunk =
3325                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3326         int dd_idx;
3327         int chunk_offset = sector_div(stripe, sectors_per_chunk);
3328         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3329
3330         raid5_compute_sector(conf,
3331                              stripe * (disks - conf->max_degraded)
3332                              *sectors_per_chunk + chunk_offset,
3333                              previous,
3334                              &dd_idx, sh);
3335 }
3336
3337 static void
3338 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3339                      struct stripe_head_state *s, int disks)
3340 {
3341         int i;
3342         BUG_ON(sh->batch_head);
3343         for (i = disks; i--; ) {
3344                 struct bio *bi;
3345                 int bitmap_end = 0;
3346
3347                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3348                         struct md_rdev *rdev;
3349                         rcu_read_lock();
3350                         rdev = rcu_dereference(conf->disks[i].rdev);
3351                         if (rdev && test_bit(In_sync, &rdev->flags) &&
3352                             !test_bit(Faulty, &rdev->flags))
3353                                 atomic_inc(&rdev->nr_pending);
3354                         else
3355                                 rdev = NULL;
3356                         rcu_read_unlock();
3357                         if (rdev) {
3358                                 if (!rdev_set_badblocks(
3359                                             rdev,
3360                                             sh->sector,
3361                                             STRIPE_SECTORS, 0))
3362                                         md_error(conf->mddev, rdev);
3363                                 rdev_dec_pending(rdev, conf->mddev);
3364                         }
3365                 }
3366                 spin_lock_irq(&sh->stripe_lock);
3367                 /* fail all writes first */
3368                 bi = sh->dev[i].towrite;
3369                 sh->dev[i].towrite = NULL;
3370                 sh->overwrite_disks = 0;
3371                 spin_unlock_irq(&sh->stripe_lock);
3372                 if (bi)
3373                         bitmap_end = 1;
3374
3375                 log_stripe_write_finished(sh);
3376
3377                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3378                         wake_up(&conf->wait_for_overlap);
3379
3380                 while (bi && bi->bi_iter.bi_sector <
3381                         sh->dev[i].sector + STRIPE_SECTORS) {
3382                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3383
3384                         md_write_end(conf->mddev);
3385                         bio_io_error(bi);
3386                         bi = nextbi;
3387                 }
3388                 if (bitmap_end)
3389                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3390                                 STRIPE_SECTORS, 0, 0);
3391                 bitmap_end = 0;
3392                 /* and fail all 'written' */
3393                 bi = sh->dev[i].written;
3394                 sh->dev[i].written = NULL;
3395                 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3396                         WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3397                         sh->dev[i].page = sh->dev[i].orig_page;
3398                 }
3399
3400                 if (bi) bitmap_end = 1;
3401                 while (bi && bi->bi_iter.bi_sector <
3402                        sh->dev[i].sector + STRIPE_SECTORS) {
3403                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3404
3405                         md_write_end(conf->mddev);
3406                         bio_io_error(bi);
3407                         bi = bi2;
3408                 }
3409
3410                 /* fail any reads if this device is non-operational and
3411                  * the data has not reached the cache yet.
3412                  */
3413                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3414                     s->failed > conf->max_degraded &&
3415                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3416                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
3417                         spin_lock_irq(&sh->stripe_lock);
3418                         bi = sh->dev[i].toread;
3419                         sh->dev[i].toread = NULL;
3420                         spin_unlock_irq(&sh->stripe_lock);
3421                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3422                                 wake_up(&conf->wait_for_overlap);
3423                         if (bi)
3424                                 s->to_read--;
3425                         while (bi && bi->bi_iter.bi_sector <
3426                                sh->dev[i].sector + STRIPE_SECTORS) {
3427                                 struct bio *nextbi =
3428                                         r5_next_bio(bi, sh->dev[i].sector);
3429
3430                                 bio_io_error(bi);
3431                                 bi = nextbi;
3432                         }
3433                 }
3434                 if (bitmap_end)
3435                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3436                                         STRIPE_SECTORS, 0, 0);
3437                 /* If we were in the middle of a write the parity block might
3438                  * still be locked - so just clear all R5_LOCKED flags
3439                  */
3440                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3441         }
3442         s->to_write = 0;
3443         s->written = 0;
3444
3445         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3446                 if (atomic_dec_and_test(&conf->pending_full_writes))
3447                         md_wakeup_thread(conf->mddev->thread);
3448 }
3449
3450 static void
3451 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3452                    struct stripe_head_state *s)
3453 {
3454         int abort = 0;
3455         int i;
3456
3457         BUG_ON(sh->batch_head);
3458         clear_bit(STRIPE_SYNCING, &sh->state);
3459         if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3460                 wake_up(&conf->wait_for_overlap);
3461         s->syncing = 0;
3462         s->replacing = 0;
3463         /* There is nothing more to do for sync/check/repair.
3464          * Don't even need to abort as that is handled elsewhere
3465          * if needed, and not always wanted e.g. if there is a known
3466          * bad block here.
3467          * For recover/replace we need to record a bad block on all
3468          * non-sync devices, or abort the recovery
3469          */
3470         if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3471                 /* During recovery devices cannot be removed, so
3472                  * locking and refcounting of rdevs is not needed
3473                  */
3474                 rcu_read_lock();
3475                 for (i = 0; i < conf->raid_disks; i++) {
3476                         struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3477                         if (rdev
3478                             && !test_bit(Faulty, &rdev->flags)
3479                             && !test_bit(In_sync, &rdev->flags)
3480                             && !rdev_set_badblocks(rdev, sh->sector,
3481                                                    STRIPE_SECTORS, 0))
3482                                 abort = 1;
3483                         rdev = rcu_dereference(conf->disks[i].replacement);
3484                         if (rdev
3485                             && !test_bit(Faulty, &rdev->flags)
3486                             && !test_bit(In_sync, &rdev->flags)
3487                             && !rdev_set_badblocks(rdev, sh->sector,
3488                                                    STRIPE_SECTORS, 0))
3489                                 abort = 1;
3490                 }
3491                 rcu_read_unlock();
3492                 if (abort)
3493                         conf->recovery_disabled =
3494                                 conf->mddev->recovery_disabled;
3495         }
3496         md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3497 }
3498
3499 static int want_replace(struct stripe_head *sh, int disk_idx)
3500 {
3501         struct md_rdev *rdev;
3502         int rv = 0;
3503
3504         rcu_read_lock();
3505         rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3506         if (rdev
3507             && !test_bit(Faulty, &rdev->flags)
3508             && !test_bit(In_sync, &rdev->flags)
3509             && (rdev->recovery_offset <= sh->sector
3510                 || rdev->mddev->recovery_cp <= sh->sector))
3511                 rv = 1;
3512         rcu_read_unlock();
3513         return rv;
3514 }
3515
3516 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3517                            int disk_idx, int disks)
3518 {
3519         struct r5dev *dev = &sh->dev[disk_idx];
3520         struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3521                                   &sh->dev[s->failed_num[1]] };
3522         int i;
3523
3524
3525         if (test_bit(R5_LOCKED, &dev->flags) ||
3526             test_bit(R5_UPTODATE, &dev->flags))
3527                 /* No point reading this as we already have it or have
3528                  * decided to get it.
3529                  */
3530                 return 0;
3531
3532         if (dev->toread ||
3533             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3534                 /* We need this block to directly satisfy a request */
3535                 return 1;
3536
3537         if (s->syncing || s->expanding ||
3538             (s->replacing && want_replace(sh, disk_idx)))
3539                 /* When syncing, or expanding we read everything.
3540                  * When replacing, we need the replaced block.
3541                  */
3542                 return 1;
3543
3544         if ((s->failed >= 1 && fdev[0]->toread) ||
3545             (s->failed >= 2 && fdev[1]->toread))
3546                 /* If we want to read from a failed device, then
3547                  * we need to actually read every other device.
3548                  */
3549                 return 1;
3550
3551         /* Sometimes neither read-modify-write nor reconstruct-write
3552          * cycles can work.  In those cases we read every block we
3553          * can.  Then the parity-update is certain to have enough to
3554          * work with.
3555          * This can only be a problem when we need to write something,
3556          * and some device has failed.  If either of those tests
3557          * fail we need look no further.
3558          */
3559         if (!s->failed || !s->to_write)
3560                 return 0;
3561
3562         if (test_bit(R5_Insync, &dev->flags) &&
3563             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3564                 /* Pre-reads at not permitted until after short delay
3565                  * to gather multiple requests.  However if this
3566                  * device is no Insync, the block could only be computed
3567                  * and there is no need to delay that.
3568                  */
3569                 return 0;
3570
3571         for (i = 0; i < s->failed && i < 2; i++) {
3572                 if (fdev[i]->towrite &&
3573                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3574                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3575                         /* If we have a partial write to a failed
3576                          * device, then we will need to reconstruct
3577                          * the content of that device, so all other
3578                          * devices must be read.
3579                          */
3580                         return 1;
3581         }
3582
3583         /* If we are forced to do a reconstruct-write, either because
3584          * the current RAID6 implementation only supports that, or
3585          * because parity cannot be trusted and we are currently
3586          * recovering it, there is extra need to be careful.
3587          * If one of the devices that we would need to read, because
3588          * it is not being overwritten (and maybe not written at all)
3589          * is missing/faulty, then we need to read everything we can.
3590          */
3591         if (sh->raid_conf->level != 6 &&
3592             sh->sector < sh->raid_conf->mddev->recovery_cp)
3593                 /* reconstruct-write isn't being forced */
3594                 return 0;
3595         for (i = 0; i < s->failed && i < 2; i++) {
3596                 if (s->failed_num[i] != sh->pd_idx &&
3597                     s->failed_num[i] != sh->qd_idx &&
3598                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3599                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3600                         return 1;
3601         }
3602
3603         return 0;
3604 }
3605
3606 /* fetch_block - checks the given member device to see if its data needs
3607  * to be read or computed to satisfy a request.
3608  *
3609  * Returns 1 when no more member devices need to be checked, otherwise returns
3610  * 0 to tell the loop in handle_stripe_fill to continue
3611  */
3612 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3613                        int disk_idx, int disks)
3614 {
3615         struct r5dev *dev = &sh->dev[disk_idx];
3616
3617         /* is the data in this block needed, and can we get it? */
3618         if (need_this_block(sh, s, disk_idx, disks)) {
3619                 /* we would like to get this block, possibly by computing it,
3620                  * otherwise read it if the backing disk is insync
3621                  */
3622                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3623                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3624                 BUG_ON(sh->batch_head);
3625
3626                 /*
3627                  * In the raid6 case if the only non-uptodate disk is P
3628                  * then we already trusted P to compute the other failed
3629                  * drives. It is safe to compute rather than re-read P.
3630                  * In other cases we only compute blocks from failed
3631                  * devices, otherwise check/repair might fail to detect
3632                  * a real inconsistency.
3633                  */
3634
3635                 if ((s->uptodate == disks - 1) &&
3636                     ((sh->qd_idx >= 0 && sh->pd_idx == disk_idx) ||
3637                     (s->failed && (disk_idx == s->failed_num[0] ||
3638                                    disk_idx == s->failed_num[1])))) {
3639                         /* have disk failed, and we're requested to fetch it;
3640                          * do compute it
3641                          */
3642                         pr_debug("Computing stripe %llu block %d\n",
3643                                (unsigned long long)sh->sector, disk_idx);
3644                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3645                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3646                         set_bit(R5_Wantcompute, &dev->flags);
3647                         sh->ops.target = disk_idx;
3648                         sh->ops.target2 = -1; /* no 2nd target */
3649                         s->req_compute = 1;
3650                         /* Careful: from this point on 'uptodate' is in the eye
3651                          * of raid_run_ops which services 'compute' operations
3652                          * before writes. R5_Wantcompute flags a block that will
3653                          * be R5_UPTODATE by the time it is needed for a
3654                          * subsequent operation.
3655                          */
3656                         s->uptodate++;
3657                         return 1;
3658                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3659                         /* Computing 2-failure is *very* expensive; only
3660                          * do it if failed >= 2
3661                          */
3662                         int other;
3663                         for (other = disks; other--; ) {
3664                                 if (other == disk_idx)
3665                                         continue;
3666                                 if (!test_bit(R5_UPTODATE,
3667                                       &sh->dev[other].flags))
3668                                         break;
3669                         }
3670                         BUG_ON(other < 0);
3671                         pr_debug("Computing stripe %llu blocks %d,%d\n",
3672                                (unsigned long long)sh->sector,
3673                                disk_idx, other);
3674                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3675                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3676                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3677                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
3678                         sh->ops.target = disk_idx;
3679                         sh->ops.target2 = other;
3680                         s->uptodate += 2;
3681                         s->req_compute = 1;
3682                         return 1;
3683                 } else if (test_bit(R5_Insync, &dev->flags)) {
3684                         set_bit(R5_LOCKED, &dev->flags);
3685                         set_bit(R5_Wantread, &dev->flags);
3686                         s->locked++;
3687                         pr_debug("Reading block %d (sync=%d)\n",
3688                                 disk_idx, s->syncing);
3689                 }
3690         }
3691
3692         return 0;
3693 }
3694
3695 /**
3696  * handle_stripe_fill - read or compute data to satisfy pending requests.
3697  */
3698 static void handle_stripe_fill(struct stripe_head *sh,
3699                                struct stripe_head_state *s,
3700                                int disks)
3701 {
3702         int i;
3703
3704         /* look for blocks to read/compute, skip this if a compute
3705          * is already in flight, or if the stripe contents are in the
3706          * midst of changing due to a write
3707          */
3708         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3709             !sh->reconstruct_state) {
3710
3711                 /*
3712                  * For degraded stripe with data in journal, do not handle
3713                  * read requests yet, instead, flush the stripe to raid
3714                  * disks first, this avoids handling complex rmw of write
3715                  * back cache (prexor with orig_page, and then xor with
3716                  * page) in the read path
3717                  */
3718                 if (s->injournal && s->failed) {
3719                         if (test_bit(STRIPE_R5C_CACHING, &sh->state))
3720                                 r5c_make_stripe_write_out(sh);
3721                         goto out;
3722                 }
3723
3724                 for (i = disks; i--; )
3725                         if (fetch_block(sh, s, i, disks))
3726                                 break;
3727         }
3728 out:
3729         set_bit(STRIPE_HANDLE, &sh->state);
3730 }
3731
3732 static void break_stripe_batch_list(struct stripe_head *head_sh,
3733                                     unsigned long handle_flags);
3734 /* handle_stripe_clean_event
3735  * any written block on an uptodate or failed drive can be returned.
3736  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3737  * never LOCKED, so we don't need to test 'failed' directly.
3738  */
3739 static void handle_stripe_clean_event(struct r5conf *conf,
3740         struct stripe_head *sh, int disks)
3741 {
3742         int i;
3743         struct r5dev *dev;
3744         int discard_pending = 0;
3745         struct stripe_head *head_sh = sh;
3746         bool do_endio = false;
3747
3748         for (i = disks; i--; )
3749                 if (sh->dev[i].written) {
3750                         dev = &sh->dev[i];
3751                         if (!test_bit(R5_LOCKED, &dev->flags) &&
3752                             (test_bit(R5_UPTODATE, &dev->flags) ||
3753                              test_bit(R5_Discard, &dev->flags) ||
3754                              test_bit(R5_SkipCopy, &dev->flags))) {
3755                                 /* We can return any write requests */
3756                                 struct bio *wbi, *wbi2;
3757                                 pr_debug("Return write for disc %d\n", i);
3758                                 if (test_and_clear_bit(R5_Discard, &dev->flags))
3759                                         clear_bit(R5_UPTODATE, &dev->flags);
3760                                 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3761                                         WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3762                                 }
3763                                 do_endio = true;
3764
3765 returnbi:
3766                                 dev->page = dev->orig_page;
3767                                 wbi = dev->written;
3768                                 dev->written = NULL;
3769                                 while (wbi && wbi->bi_iter.bi_sector <
3770                                         dev->sector + STRIPE_SECTORS) {
3771                                         wbi2 = r5_next_bio(wbi, dev->sector);
3772                                         md_write_end(conf->mddev);
3773                                         bio_endio(wbi);
3774                                         wbi = wbi2;
3775                                 }
3776                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3777                                                 STRIPE_SECTORS,
3778                                          !test_bit(STRIPE_DEGRADED, &sh->state),
3779                                                 0);
3780                                 if (head_sh->batch_head) {
3781                                         sh = list_first_entry(&sh->batch_list,
3782                                                               struct stripe_head,
3783                                                               batch_list);
3784                                         if (sh != head_sh) {
3785                                                 dev = &sh->dev[i];
3786                                                 goto returnbi;
3787                                         }
3788                                 }
3789                                 sh = head_sh;
3790                                 dev = &sh->dev[i];
3791                         } else if (test_bit(R5_Discard, &dev->flags))
3792                                 discard_pending = 1;
3793                 }
3794
3795         log_stripe_write_finished(sh);
3796
3797         if (!discard_pending &&
3798             test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3799                 int hash;
3800                 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3801                 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3802                 if (sh->qd_idx >= 0) {
3803                         clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3804                         clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3805                 }
3806                 /* now that discard is done we can proceed with any sync */
3807                 clear_bit(STRIPE_DISCARD, &sh->state);
3808                 /*
3809                  * SCSI discard will change some bio fields and the stripe has
3810                  * no updated data, so remove it from hash list and the stripe
3811                  * will be reinitialized
3812                  */
3813 unhash:
3814                 hash = sh->hash_lock_index;
3815                 spin_lock_irq(conf->hash_locks + hash);
3816                 remove_hash(sh);
3817                 spin_unlock_irq(conf->hash_locks + hash);
3818                 if (head_sh->batch_head) {
3819                         sh = list_first_entry(&sh->batch_list,
3820                                               struct stripe_head, batch_list);
3821                         if (sh != head_sh)
3822                                         goto unhash;
3823                 }
3824                 sh = head_sh;
3825
3826                 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3827                         set_bit(STRIPE_HANDLE, &sh->state);
3828
3829         }
3830
3831         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3832                 if (atomic_dec_and_test(&conf->pending_full_writes))
3833                         md_wakeup_thread(conf->mddev->thread);
3834
3835         if (head_sh->batch_head && do_endio)
3836                 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3837 }
3838
3839 /*
3840  * For RMW in write back cache, we need extra page in prexor to store the
3841  * old data. This page is stored in dev->orig_page.
3842  *
3843  * This function checks whether we have data for prexor. The exact logic
3844  * is:
3845  *       R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
3846  */
3847 static inline bool uptodate_for_rmw(struct r5dev *dev)
3848 {
3849         return (test_bit(R5_UPTODATE, &dev->flags)) &&
3850                 (!test_bit(R5_InJournal, &dev->flags) ||
3851                  test_bit(R5_OrigPageUPTDODATE, &dev->flags));
3852 }
3853
3854 static int handle_stripe_dirtying(struct r5conf *conf,
3855                                   struct stripe_head *sh,
3856                                   struct stripe_head_state *s,
3857                                   int disks)
3858 {
3859         int rmw = 0, rcw = 0, i;
3860         sector_t recovery_cp = conf->mddev->recovery_cp;
3861
3862         /* Check whether resync is now happening or should start.
3863          * If yes, then the array is dirty (after unclean shutdown or
3864          * initial creation), so parity in some stripes might be inconsistent.
3865          * In this case, we need to always do reconstruct-write, to ensure
3866          * that in case of drive failure or read-error correction, we
3867          * generate correct data from the parity.
3868          */
3869         if (conf->rmw_level == PARITY_DISABLE_RMW ||
3870             (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3871              s->failed == 0)) {
3872                 /* Calculate the real rcw later - for now make it
3873                  * look like rcw is cheaper
3874                  */
3875                 rcw = 1; rmw = 2;
3876                 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3877                          conf->rmw_level, (unsigned long long)recovery_cp,
3878                          (unsigned long long)sh->sector);
3879         } else for (i = disks; i--; ) {
3880                 /* would I have to read this buffer for read_modify_write */
3881                 struct r5dev *dev = &sh->dev[i];
3882                 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3883                      i == sh->pd_idx || i == sh->qd_idx ||
3884                      test_bit(R5_InJournal, &dev->flags)) &&
3885                     !test_bit(R5_LOCKED, &dev->flags) &&
3886                     !(uptodate_for_rmw(dev) ||
3887                       test_bit(R5_Wantcompute, &dev->flags))) {
3888                         if (test_bit(R5_Insync, &dev->flags))
3889                                 rmw++;
3890                         else
3891                                 rmw += 2*disks;  /* cannot read it */
3892                 }
3893                 /* Would I have to read this buffer for reconstruct_write */
3894                 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3895                     i != sh->pd_idx && i != sh->qd_idx &&
3896                     !test_bit(R5_LOCKED, &dev->flags) &&
3897                     !(test_bit(R5_UPTODATE, &dev->flags) ||
3898                       test_bit(R5_Wantcompute, &dev->flags))) {
3899                         if (test_bit(R5_Insync, &dev->flags))
3900                                 rcw++;
3901                         else
3902                                 rcw += 2*disks;
3903                 }
3904         }
3905
3906         pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
3907                  (unsigned long long)sh->sector, sh->state, rmw, rcw);
3908         set_bit(STRIPE_HANDLE, &sh->state);
3909         if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
3910                 /* prefer read-modify-write, but need to get some data */
3911                 if (conf->mddev->queue)
3912                         blk_add_trace_msg(conf->mddev->queue,
3913                                           "raid5 rmw %llu %d",
3914                                           (unsigned long long)sh->sector, rmw);
3915                 for (i = disks; i--; ) {
3916                         struct r5dev *dev = &sh->dev[i];
3917                         if (test_bit(R5_InJournal, &dev->flags) &&
3918                             dev->page == dev->orig_page &&
3919                             !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
3920                                 /* alloc page for prexor */
3921                                 struct page *p = alloc_page(GFP_NOIO);
3922
3923                                 if (p) {
3924                                         dev->orig_page = p;
3925                                         continue;
3926                                 }
3927
3928                                 /*
3929                                  * alloc_page() failed, try use
3930                                  * disk_info->extra_page
3931                                  */
3932                                 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
3933                                                       &conf->cache_state)) {
3934                                         r5c_use_extra_page(sh);
3935                                         break;
3936                                 }
3937
3938                                 /* extra_page in use, add to delayed_list */
3939                                 set_bit(STRIPE_DELAYED, &sh->state);
3940                                 s->waiting_extra_page = 1;
3941                                 return -EAGAIN;
3942                         }
3943                 }
3944
3945                 for (i = disks; i--; ) {
3946                         struct r5dev *dev = &sh->dev[i];
3947                         if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3948                              i == sh->pd_idx || i == sh->qd_idx ||
3949                              test_bit(R5_InJournal, &dev->flags)) &&
3950                             !test_bit(R5_LOCKED, &dev->flags) &&
3951                             !(uptodate_for_rmw(dev) ||
3952                               test_bit(R5_Wantcompute, &dev->flags)) &&
3953                             test_bit(R5_Insync, &dev->flags)) {
3954                                 if (test_bit(STRIPE_PREREAD_ACTIVE,
3955                                              &sh->state)) {
3956                                         pr_debug("Read_old block %d for r-m-w\n",
3957                                                  i);
3958                                         set_bit(R5_LOCKED, &dev->flags);
3959                                         set_bit(R5_Wantread, &dev->flags);
3960                                         s->locked++;
3961                                 } else {
3962                                         set_bit(STRIPE_DELAYED, &sh->state);
3963                                         set_bit(STRIPE_HANDLE, &sh->state);
3964                                 }
3965                         }
3966                 }
3967         }
3968         if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
3969                 /* want reconstruct write, but need to get some data */
3970                 int qread =0;
3971                 rcw = 0;
3972                 for (i = disks; i--; ) {
3973                         struct r5dev *dev = &sh->dev[i];
3974                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3975                             i != sh->pd_idx && i != sh->qd_idx &&
3976                             !test_bit(R5_LOCKED, &dev->flags) &&
3977                             !(test_bit(R5_UPTODATE, &dev->flags) ||
3978                               test_bit(R5_Wantcompute, &dev->flags))) {
3979                                 rcw++;
3980                                 if (test_bit(R5_Insync, &dev->flags) &&
3981                                     test_bit(STRIPE_PREREAD_ACTIVE,
3982                                              &sh->state)) {
3983                                         pr_debug("Read_old block "
3984                                                 "%d for Reconstruct\n", i);
3985                                         set_bit(R5_LOCKED, &dev->flags);
3986                                         set_bit(R5_Wantread, &dev->flags);
3987                                         s->locked++;
3988                                         qread++;
3989                                 } else {
3990                                         set_bit(STRIPE_DELAYED, &sh->state);
3991                                         set_bit(STRIPE_HANDLE, &sh->state);
3992                                 }
3993                         }
3994                 }
3995                 if (rcw && conf->mddev->queue)
3996                         blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3997                                           (unsigned long long)sh->sector,
3998                                           rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3999         }
4000
4001         if (rcw > disks && rmw > disks &&
4002             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4003                 set_bit(STRIPE_DELAYED, &sh->state);
4004
4005         /* now if nothing is locked, and if we have enough data,
4006          * we can start a write request
4007          */
4008         /* since handle_stripe can be called at any time we need to handle the
4009          * case where a compute block operation has been submitted and then a
4010          * subsequent call wants to start a write request.  raid_run_ops only
4011          * handles the case where compute block and reconstruct are requested
4012          * simultaneously.  If this is not the case then new writes need to be
4013          * held off until the compute completes.
4014          */
4015         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
4016             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
4017              !test_bit(STRIPE_BIT_DELAY, &sh->state)))
4018                 schedule_reconstruction(sh, s, rcw == 0, 0);
4019         return 0;
4020 }
4021
4022 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
4023                                 struct stripe_head_state *s, int disks)
4024 {
4025         struct r5dev *dev = NULL;
4026
4027         BUG_ON(sh->batch_head);
4028         set_bit(STRIPE_HANDLE, &sh->state);
4029
4030         switch (sh->check_state) {
4031         case check_state_idle:
4032                 /* start a new check operation if there are no failures */
4033                 if (s->failed == 0) {
4034                         BUG_ON(s->uptodate != disks);
4035                         sh->check_state = check_state_run;
4036                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
4037                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4038                         s->uptodate--;
4039                         break;
4040                 }
4041                 dev = &sh->dev[s->failed_num[0]];
4042                 /* fall through */
4043         case check_state_compute_result:
4044                 sh->check_state = check_state_idle;
4045                 if (!dev)
4046                         dev = &sh->dev[sh->pd_idx];
4047
4048                 /* check that a write has not made the stripe insync */
4049                 if (test_bit(STRIPE_INSYNC, &sh->state))
4050                         break;
4051
4052                 /* either failed parity check, or recovery is happening */
4053                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
4054                 BUG_ON(s->uptodate != disks);
4055
4056                 set_bit(R5_LOCKED, &dev->flags);
4057                 s->locked++;
4058                 set_bit(R5_Wantwrite, &dev->flags);
4059
4060                 clear_bit(STRIPE_DEGRADED, &sh->state);
4061                 set_bit(STRIPE_INSYNC, &sh->state);
4062                 break;
4063         case check_state_run:
4064                 break; /* we will be called again upon completion */
4065         case check_state_check_result:
4066                 sh->check_state = check_state_idle;
4067
4068                 /* if a failure occurred during the check operation, leave
4069                  * STRIPE_INSYNC not set and let the stripe be handled again
4070                  */
4071                 if (s->failed)
4072                         break;
4073
4074                 /* handle a successful check operation, if parity is correct
4075                  * we are done.  Otherwise update the mismatch count and repair
4076                  * parity if !MD_RECOVERY_CHECK
4077                  */
4078                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
4079                         /* parity is correct (on disc,
4080                          * not in buffer any more)
4081                          */
4082                         set_bit(STRIPE_INSYNC, &sh->state);
4083                 else {
4084                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4085                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4086                                 /* don't try to repair!! */
4087                                 set_bit(STRIPE_INSYNC, &sh->state);
4088                                 pr_warn_ratelimited("%s: mismatch sector in range "
4089                                                     "%llu-%llu\n", mdname(conf->mddev),
4090                                                     (unsigned long long) sh->sector,
4091                                                     (unsigned long long) sh->sector +
4092                                                     STRIPE_SECTORS);
4093                         } else {
4094                                 sh->check_state = check_state_compute_run;
4095                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4096                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4097                                 set_bit(R5_Wantcompute,
4098                                         &sh->dev[sh->pd_idx].flags);
4099                                 sh->ops.target = sh->pd_idx;
4100                                 sh->ops.target2 = -1;
4101                                 s->uptodate++;
4102                         }
4103                 }
4104                 break;
4105         case check_state_compute_run:
4106                 break;
4107         default:
4108                 pr_err("%s: unknown check_state: %d sector: %llu\n",
4109                        __func__, sh->check_state,
4110                        (unsigned long long) sh->sector);
4111                 BUG();
4112         }
4113 }
4114
4115 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
4116                                   struct stripe_head_state *s,
4117                                   int disks)
4118 {
4119         int pd_idx = sh->pd_idx;
4120         int qd_idx = sh->qd_idx;
4121         struct r5dev *dev;
4122
4123         BUG_ON(sh->batch_head);
4124         set_bit(STRIPE_HANDLE, &sh->state);
4125
4126         BUG_ON(s->failed > 2);
4127
4128         /* Want to check and possibly repair P and Q.
4129          * However there could be one 'failed' device, in which
4130          * case we can only check one of them, possibly using the
4131          * other to generate missing data
4132          */
4133
4134         switch (sh->check_state) {
4135         case check_state_idle:
4136                 /* start a new check operation if there are < 2 failures */
4137                 if (s->failed == s->q_failed) {
4138                         /* The only possible failed device holds Q, so it
4139                          * makes sense to check P (If anything else were failed,
4140                          * we would have used P to recreate it).
4141                          */
4142                         sh->check_state = check_state_run;
4143                 }
4144                 if (!s->q_failed && s->failed < 2) {
4145                         /* Q is not failed, and we didn't use it to generate
4146                          * anything, so it makes sense to check it
4147                          */
4148                         if (sh->check_state == check_state_run)
4149                                 sh->check_state = check_state_run_pq;
4150                         else
4151                                 sh->check_state = check_state_run_q;
4152                 }
4153
4154                 /* discard potentially stale zero_sum_result */
4155                 sh->ops.zero_sum_result = 0;
4156
4157                 if (sh->check_state == check_state_run) {
4158                         /* async_xor_zero_sum destroys the contents of P */
4159                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
4160                         s->uptodate--;
4161                 }
4162                 if (sh->check_state >= check_state_run &&
4163                     sh->check_state <= check_state_run_pq) {
4164                         /* async_syndrome_zero_sum preserves P and Q, so
4165                          * no need to mark them !uptodate here
4166                          */
4167                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
4168                         break;
4169                 }
4170
4171                 /* we have 2-disk failure */
4172                 BUG_ON(s->failed != 2);
4173                 /* fall through */
4174         case check_state_compute_result:
4175                 sh->check_state = check_state_idle;
4176
4177                 /* check that a write has not made the stripe insync */
4178                 if (test_bit(STRIPE_INSYNC, &sh->state))
4179                         break;
4180
4181                 /* now write out any block on a failed drive,
4182                  * or P or Q if they were recomputed
4183                  */
4184                 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
4185                 if (s->failed == 2) {
4186                         dev = &sh->dev[s->failed_num[1]];
4187                         s->locked++;
4188                         set_bit(R5_LOCKED, &dev->flags);
4189                         set_bit(R5_Wantwrite, &dev->flags);
4190                 }
4191                 if (s->failed >= 1) {
4192                         dev = &sh->dev[s->failed_num[0]];
4193                         s->locked++;
4194                         set_bit(R5_LOCKED, &dev->flags);
4195                         set_bit(R5_Wantwrite, &dev->flags);
4196                 }
4197                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4198                         dev = &sh->dev[pd_idx];
4199                         s->locked++;
4200                         set_bit(R5_LOCKED, &dev->flags);
4201                         set_bit(R5_Wantwrite, &dev->flags);
4202                 }
4203                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4204                         dev = &sh->dev[qd_idx];
4205                         s->locked++;
4206                         set_bit(R5_LOCKED, &dev->flags);
4207                         set_bit(R5_Wantwrite, &dev->flags);
4208                 }
4209                 clear_bit(STRIPE_DEGRADED, &sh->state);
4210
4211                 set_bit(STRIPE_INSYNC, &sh->state);
4212                 break;
4213         case check_state_run:
4214         case check_state_run_q:
4215         case check_state_run_pq:
4216                 break; /* we will be called again upon completion */
4217         case check_state_check_result:
4218                 sh->check_state = check_state_idle;
4219
4220                 /* handle a successful check operation, if parity is correct
4221                  * we are done.  Otherwise update the mismatch count and repair
4222                  * parity if !MD_RECOVERY_CHECK
4223                  */
4224                 if (sh->ops.zero_sum_result == 0) {
4225                         /* both parities are correct */
4226                         if (!s->failed)
4227                                 set_bit(STRIPE_INSYNC, &sh->state);
4228                         else {
4229                                 /* in contrast to the raid5 case we can validate
4230                                  * parity, but still have a failure to write
4231                                  * back
4232                                  */
4233                                 sh->check_state = check_state_compute_result;
4234                                 /* Returning at this point means that we may go
4235                                  * off and bring p and/or q uptodate again so
4236                                  * we make sure to check zero_sum_result again
4237                                  * to verify if p or q need writeback
4238                                  */
4239                         }
4240                 } else {
4241                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4242                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4243                                 /* don't try to repair!! */
4244                                 set_bit(STRIPE_INSYNC, &sh->state);
4245                                 pr_warn_ratelimited("%s: mismatch sector in range "
4246                                                     "%llu-%llu\n", mdname(conf->mddev),
4247                                                     (unsigned long long) sh->sector,
4248                                                     (unsigned long long) sh->sector +
4249                                                     STRIPE_SECTORS);
4250                         } else {
4251                                 int *target = &sh->ops.target;
4252
4253                                 sh->ops.target = -1;
4254                                 sh->ops.target2 = -1;
4255                                 sh->check_state = check_state_compute_run;
4256                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4257                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4258                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4259                                         set_bit(R5_Wantcompute,
4260                                                 &sh->dev[pd_idx].flags);
4261                                         *target = pd_idx;
4262                                         target = &sh->ops.target2;
4263                                         s->uptodate++;
4264                                 }
4265                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4266                                         set_bit(R5_Wantcompute,
4267                                                 &sh->dev[qd_idx].flags);
4268                                         *target = qd_idx;
4269                                         s->uptodate++;
4270                                 }
4271                         }
4272                 }
4273                 break;
4274         case check_state_compute_run:
4275                 break;
4276         default:
4277                 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4278                         __func__, sh->check_state,
4279                         (unsigned long long) sh->sector);
4280                 BUG();
4281         }
4282 }
4283
4284 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4285 {
4286         int i;
4287
4288         /* We have read all the blocks in this stripe and now we need to
4289          * copy some of them into a target stripe for expand.
4290          */
4291         struct dma_async_tx_descriptor *tx = NULL;
4292         BUG_ON(sh->batch_head);
4293         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4294         for (i = 0; i < sh->disks; i++)
4295                 if (i != sh->pd_idx && i != sh->qd_idx) {
4296                         int dd_idx, j;
4297                         struct stripe_head *sh2;
4298                         struct async_submit_ctl submit;
4299
4300                         sector_t bn = raid5_compute_blocknr(sh, i, 1);
4301                         sector_t s = raid5_compute_sector(conf, bn, 0,
4302                                                           &dd_idx, NULL);
4303                         sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
4304                         if (sh2 == NULL)
4305                                 /* so far only the early blocks of this stripe
4306                                  * have been requested.  When later blocks
4307                                  * get requested, we will try again
4308                                  */
4309                                 continue;
4310                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4311                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4312                                 /* must have already done this block */
4313                                 raid5_release_stripe(sh2);
4314                                 continue;
4315                         }
4316
4317                         /* place all the copies on one channel */
4318                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4319                         tx = async_memcpy(sh2->dev[dd_idx].page,
4320                                           sh->dev[i].page, 0, 0, STRIPE_SIZE,
4321                                           &submit);
4322
4323                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4324                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4325                         for (j = 0; j < conf->raid_disks; j++)
4326                                 if (j != sh2->pd_idx &&
4327                                     j != sh2->qd_idx &&
4328                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
4329                                         break;
4330                         if (j == conf->raid_disks) {
4331                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
4332                                 set_bit(STRIPE_HANDLE, &sh2->state);
4333                         }
4334                         raid5_release_stripe(sh2);
4335
4336                 }
4337         /* done submitting copies, wait for them to complete */
4338         async_tx_quiesce(&tx);
4339 }
4340
4341 /*
4342  * handle_stripe - do things to a stripe.
4343  *
4344  * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4345  * state of various bits to see what needs to be done.
4346  * Possible results:
4347  *    return some read requests which now have data
4348  *    return some write requests which are safely on storage
4349  *    schedule a read on some buffers
4350  *    schedule a write of some buffers
4351  *    return confirmation of parity correctness
4352  *
4353  */
4354
4355 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4356 {
4357         struct r5conf *conf = sh->raid_conf;
4358         int disks = sh->disks;
4359         struct r5dev *dev;
4360         int i;
4361         int do_recovery = 0;
4362
4363         memset(s, 0, sizeof(*s));
4364
4365         s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4366         s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4367         s->failed_num[0] = -1;
4368         s->failed_num[1] = -1;
4369         s->log_failed = r5l_log_disk_error(conf);
4370
4371         /* Now to look around and see what can be done */
4372         rcu_read_lock();
4373         for (i=disks; i--; ) {
4374                 struct md_rdev *rdev;
4375                 sector_t first_bad;
4376                 int bad_sectors;
4377                 int is_bad = 0;
4378
4379                 dev = &sh->dev[i];
4380
4381                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4382                          i, dev->flags,
4383                          dev->toread, dev->towrite, dev->written);
4384                 /* maybe we can reply to a read
4385                  *
4386                  * new wantfill requests are only permitted while
4387                  * ops_complete_biofill is guaranteed to be inactive
4388                  */
4389                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4390                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4391                         set_bit(R5_Wantfill, &dev->flags);
4392
4393                 /* now count some things */
4394                 if (test_bit(R5_LOCKED, &dev->flags))
4395                         s->locked++;
4396                 if (test_bit(R5_UPTODATE, &dev->flags))
4397                         s->uptodate++;
4398                 if (test_bit(R5_Wantcompute, &dev->flags)) {
4399                         s->compute++;
4400                         BUG_ON(s->compute > 2);
4401                 }
4402
4403                 if (test_bit(R5_Wantfill, &dev->flags))
4404                         s->to_fill++;
4405                 else if (dev->toread)
4406                         s->to_read++;
4407                 if (dev->towrite) {
4408                         s->to_write++;
4409                         if (!test_bit(R5_OVERWRITE, &dev->flags))
4410                                 s->non_overwrite++;
4411                 }
4412                 if (dev->written)
4413                         s->written++;
4414                 /* Prefer to use the replacement for reads, but only
4415                  * if it is recovered enough and has no bad blocks.
4416                  */
4417                 rdev = rcu_dereference(conf->disks[i].replacement);
4418                 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4419                     rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4420                     !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4421                                  &first_bad, &bad_sectors))
4422                         set_bit(R5_ReadRepl, &dev->flags);
4423                 else {
4424                         if (rdev && !test_bit(Faulty, &rdev->flags))
4425                                 set_bit(R5_NeedReplace, &dev->flags);
4426                         else
4427                                 clear_bit(R5_NeedReplace, &dev->flags);
4428                         rdev = rcu_dereference(conf->disks[i].rdev);
4429                         clear_bit(R5_ReadRepl, &dev->flags);
4430                 }
4431                 if (rdev && test_bit(Faulty, &rdev->flags))
4432                         rdev = NULL;
4433                 if (rdev) {
4434                         is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4435                                              &first_bad, &bad_sectors);
4436                         if (s->blocked_rdev == NULL
4437                             && (test_bit(Blocked, &rdev->flags)
4438                                 || is_bad < 0)) {
4439                                 if (is_bad < 0)
4440                                         set_bit(BlockedBadBlocks,
4441                                                 &rdev->flags);
4442                                 s->blocked_rdev = rdev;
4443                                 atomic_inc(&rdev->nr_pending);
4444                         }
4445                 }
4446                 clear_bit(R5_Insync, &dev->flags);
4447                 if (!rdev)
4448                         /* Not in-sync */;
4449                 else if (is_bad) {
4450                         /* also not in-sync */
4451                         if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4452                             test_bit(R5_UPTODATE, &dev->flags)) {
4453                                 /* treat as in-sync, but with a read error
4454                                  * which we can now try to correct
4455                                  */
4456                                 set_bit(R5_Insync, &dev->flags);
4457                                 set_bit(R5_ReadError, &dev->flags);
4458                         }
4459                 } else if (test_bit(In_sync, &rdev->flags))
4460                         set_bit(R5_Insync, &dev->flags);
4461                 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4462                         /* in sync if before recovery_offset */
4463                         set_bit(R5_Insync, &dev->flags);
4464                 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4465                          test_bit(R5_Expanded, &dev->flags))
4466                         /* If we've reshaped into here, we assume it is Insync.
4467                          * We will shortly update recovery_offset to make
4468                          * it official.
4469                          */
4470                         set_bit(R5_Insync, &dev->flags);
4471
4472                 if (test_bit(R5_WriteError, &dev->flags)) {
4473                         /* This flag does not apply to '.replacement'
4474                          * only to .rdev, so make sure to check that*/
4475                         struct md_rdev *rdev2 = rcu_dereference(
4476                                 conf->disks[i].rdev);
4477                         if (rdev2 == rdev)
4478                                 clear_bit(R5_Insync, &dev->flags);
4479                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4480                                 s->handle_bad_blocks = 1;
4481                                 atomic_inc(&rdev2->nr_pending);
4482                         } else
4483                                 clear_bit(R5_WriteError, &dev->flags);
4484                 }
4485                 if (test_bit(R5_MadeGood, &dev->flags)) {
4486                         /* This flag does not apply to '.replacement'
4487                          * only to .rdev, so make sure to check that*/
4488                         struct md_rdev *rdev2 = rcu_dereference(
4489                                 conf->disks[i].rdev);
4490                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4491                                 s->handle_bad_blocks = 1;
4492                                 atomic_inc(&rdev2->nr_pending);
4493                         } else
4494                                 clear_bit(R5_MadeGood, &dev->flags);
4495                 }
4496                 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4497                         struct md_rdev *rdev2 = rcu_dereference(
4498                                 conf->disks[i].replacement);
4499                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4500                                 s->handle_bad_blocks = 1;
4501                                 atomic_inc(&rdev2->nr_pending);
4502                         } else
4503                                 clear_bit(R5_MadeGoodRepl, &dev->flags);
4504                 }
4505                 if (!test_bit(R5_Insync, &dev->flags)) {
4506                         /* The ReadError flag will just be confusing now */
4507                         clear_bit(R5_ReadError, &dev->flags);
4508                         clear_bit(R5_ReWrite, &dev->flags);
4509                 }
4510                 if (test_bit(R5_ReadError, &dev->flags))
4511                         clear_bit(R5_Insync, &dev->flags);
4512                 if (!test_bit(R5_Insync, &dev->flags)) {
4513                         if (s->failed < 2)
4514                                 s->failed_num[s->failed] = i;
4515                         s->failed++;
4516                         if (rdev && !test_bit(Faulty, &rdev->flags))
4517                                 do_recovery = 1;
4518                 }
4519
4520                 if (test_bit(R5_InJournal, &dev->flags))
4521                         s->injournal++;
4522                 if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4523                         s->just_cached++;
4524         }
4525         if (test_bit(STRIPE_SYNCING, &sh->state)) {
4526                 /* If there is a failed device being replaced,
4527                  *     we must be recovering.
4528                  * else if we are after recovery_cp, we must be syncing
4529                  * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4530                  * else we can only be replacing
4531                  * sync and recovery both need to read all devices, and so
4532                  * use the same flag.
4533                  */
4534                 if (do_recovery ||
4535                     sh->sector >= conf->mddev->recovery_cp ||
4536                     test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4537                         s->syncing = 1;
4538                 else
4539                         s->replacing = 1;
4540         }
4541         rcu_read_unlock();
4542 }
4543
4544 static int clear_batch_ready(struct stripe_head *sh)
4545 {
4546         /* Return '1' if this is a member of batch, or
4547          * '0' if it is a lone stripe or a head which can now be
4548          * handled.
4549          */
4550         struct stripe_head *tmp;
4551         if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4552                 return (sh->batch_head && sh->batch_head != sh);
4553         spin_lock(&sh->stripe_lock);
4554         if (!sh->batch_head) {
4555                 spin_unlock(&sh->stripe_lock);
4556                 return 0;
4557         }
4558
4559         /*
4560          * this stripe could be added to a batch list before we check
4561          * BATCH_READY, skips it
4562          */
4563         if (sh->batch_head != sh) {
4564                 spin_unlock(&sh->stripe_lock);
4565                 return 1;
4566         }
4567         spin_lock(&sh->batch_lock);
4568         list_for_each_entry(tmp, &sh->batch_list, batch_list)
4569                 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4570         spin_unlock(&sh->batch_lock);
4571         spin_unlock(&sh->stripe_lock);
4572
4573         /*
4574          * BATCH_READY is cleared, no new stripes can be added.
4575          * batch_list can be accessed without lock
4576          */
4577         return 0;
4578 }
4579
4580 static void break_stripe_batch_list(struct stripe_head *head_sh,
4581                                     unsigned long handle_flags)
4582 {
4583         struct stripe_head *sh, *next;
4584         int i;
4585         int do_wakeup = 0;
4586
4587         list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4588
4589                 list_del_init(&sh->batch_list);
4590
4591                 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4592                                           (1 << STRIPE_SYNCING) |
4593                                           (1 << STRIPE_REPLACED) |
4594                                           (1 << STRIPE_DELAYED) |
4595                                           (1 << STRIPE_BIT_DELAY) |
4596                                           (1 << STRIPE_FULL_WRITE) |
4597                                           (1 << STRIPE_BIOFILL_RUN) |
4598                                           (1 << STRIPE_COMPUTE_RUN)  |
4599                                           (1 << STRIPE_OPS_REQ_PENDING) |
4600                                           (1 << STRIPE_DISCARD) |
4601                                           (1 << STRIPE_BATCH_READY) |
4602                                           (1 << STRIPE_BATCH_ERR) |
4603                                           (1 << STRIPE_BITMAP_PENDING)),
4604                         "stripe state: %lx\n", sh->state);
4605                 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4606                                               (1 << STRIPE_REPLACED)),
4607                         "head stripe state: %lx\n", head_sh->state);
4608
4609                 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4610                                             (1 << STRIPE_PREREAD_ACTIVE) |
4611                                             (1 << STRIPE_DEGRADED) |
4612                                             (1 << STRIPE_ON_UNPLUG_LIST)),
4613                               head_sh->state & (1 << STRIPE_INSYNC));
4614
4615                 sh->check_state = head_sh->check_state;
4616                 sh->reconstruct_state = head_sh->reconstruct_state;
4617                 for (i = 0; i < sh->disks; i++) {
4618                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4619                                 do_wakeup = 1;
4620                         sh->dev[i].flags = head_sh->dev[i].flags &
4621                                 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4622                 }
4623                 spin_lock_irq(&sh->stripe_lock);
4624                 sh->batch_head = NULL;
4625                 spin_unlock_irq(&sh->stripe_lock);
4626                 if (handle_flags == 0 ||
4627                     sh->state & handle_flags)
4628                         set_bit(STRIPE_HANDLE, &sh->state);
4629                 raid5_release_stripe(sh);
4630         }
4631         spin_lock_irq(&head_sh->stripe_lock);
4632         head_sh->batch_head = NULL;
4633         spin_unlock_irq(&head_sh->stripe_lock);
4634         for (i = 0; i < head_sh->disks; i++)
4635                 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4636                         do_wakeup = 1;
4637         if (head_sh->state & handle_flags)
4638                 set_bit(STRIPE_HANDLE, &head_sh->state);
4639
4640         if (do_wakeup)
4641                 wake_up(&head_sh->raid_conf->wait_for_overlap);
4642 }
4643
4644 static void handle_stripe(struct stripe_head *sh)
4645 {
4646         struct stripe_head_state s;
4647         struct r5conf *conf = sh->raid_conf;
4648         int i;
4649         int prexor;
4650         int disks = sh->disks;
4651         struct r5dev *pdev, *qdev;
4652
4653         clear_bit(STRIPE_HANDLE, &sh->state);
4654         if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4655                 /* already being handled, ensure it gets handled
4656                  * again when current action finishes */
4657                 set_bit(STRIPE_HANDLE, &sh->state);
4658                 return;
4659         }
4660
4661         if (clear_batch_ready(sh) ) {
4662                 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4663                 return;
4664         }
4665
4666         if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4667                 break_stripe_batch_list(sh, 0);
4668
4669         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4670                 spin_lock(&sh->stripe_lock);
4671                 /*
4672                  * Cannot process 'sync' concurrently with 'discard'.
4673                  * Flush data in r5cache before 'sync'.
4674                  */
4675                 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
4676                     !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) &&
4677                     !test_bit(STRIPE_DISCARD, &sh->state) &&
4678                     test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4679                         set_bit(STRIPE_SYNCING, &sh->state);
4680                         clear_bit(STRIPE_INSYNC, &sh->state);
4681                         clear_bit(STRIPE_REPLACED, &sh->state);
4682                 }
4683                 spin_unlock(&sh->stripe_lock);
4684         }
4685         clear_bit(STRIPE_DELAYED, &sh->state);
4686
4687         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4688                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4689                (unsigned long long)sh->sector, sh->state,
4690                atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4691                sh->check_state, sh->reconstruct_state);
4692
4693         analyse_stripe(sh, &s);
4694
4695         if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4696                 goto finish;
4697
4698         if (s.handle_bad_blocks ||
4699             test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
4700                 set_bit(STRIPE_HANDLE, &sh->state);
4701                 goto finish;
4702         }
4703
4704         if (unlikely(s.blocked_rdev)) {
4705                 if (s.syncing || s.expanding || s.expanded ||
4706                     s.replacing || s.to_write || s.written) {
4707                         set_bit(STRIPE_HANDLE, &sh->state);
4708                         goto finish;
4709                 }
4710                 /* There is nothing for the blocked_rdev to block */
4711                 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4712                 s.blocked_rdev = NULL;
4713         }
4714
4715         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4716                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4717                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4718         }
4719
4720         pr_debug("locked=%d uptodate=%d to_read=%d"
4721                " to_write=%d failed=%d failed_num=%d,%d\n",
4722                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4723                s.failed_num[0], s.failed_num[1]);
4724         /*
4725          * check if the array has lost more than max_degraded devices and,
4726          * if so, some requests might need to be failed.
4727          *
4728          * When journal device failed (log_failed), we will only process
4729          * the stripe if there is data need write to raid disks
4730          */
4731         if (s.failed > conf->max_degraded ||
4732             (s.log_failed && s.injournal == 0)) {
4733                 sh->check_state = 0;
4734                 sh->reconstruct_state = 0;
4735                 break_stripe_batch_list(sh, 0);
4736                 if (s.to_read+s.to_write+s.written)
4737                         handle_failed_stripe(conf, sh, &s, disks);
4738                 if (s.syncing + s.replacing)
4739                         handle_failed_sync(conf, sh, &s);
4740         }
4741
4742         /* Now we check to see if any write operations have recently
4743          * completed
4744          */
4745         prexor = 0;
4746         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4747                 prexor = 1;
4748         if (sh->reconstruct_state == reconstruct_state_drain_result ||
4749             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4750                 sh->reconstruct_state = reconstruct_state_idle;
4751
4752                 /* All the 'written' buffers and the parity block are ready to
4753                  * be written back to disk
4754                  */
4755                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4756                        !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4757                 BUG_ON(sh->qd_idx >= 0 &&
4758                        !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4759                        !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4760                 for (i = disks; i--; ) {
4761                         struct r5dev *dev = &sh->dev[i];
4762                         if (test_bit(R5_LOCKED, &dev->flags) &&
4763                                 (i == sh->pd_idx || i == sh->qd_idx ||
4764                                  dev->written || test_bit(R5_InJournal,
4765                                                           &dev->flags))) {
4766                                 pr_debug("Writing block %d\n", i);
4767                                 set_bit(R5_Wantwrite, &dev->flags);
4768                                 if (prexor)
4769                                         continue;
4770                                 if (s.failed > 1)
4771                                         continue;
4772                                 if (!test_bit(R5_Insync, &dev->flags) ||
4773                                     ((i == sh->pd_idx || i == sh->qd_idx)  &&
4774                                      s.failed == 0))
4775                                         set_bit(STRIPE_INSYNC, &sh->state);
4776                         }
4777                 }
4778                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4779                         s.dec_preread_active = 1;
4780         }
4781
4782         /*
4783          * might be able to return some write requests if the parity blocks
4784          * are safe, or on a failed drive
4785          */
4786         pdev = &sh->dev[sh->pd_idx];
4787         s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4788                 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4789         qdev = &sh->dev[sh->qd_idx];
4790         s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4791                 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4792                 || conf->level < 6;
4793
4794         if (s.written &&
4795             (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4796                              && !test_bit(R5_LOCKED, &pdev->flags)
4797                              && (test_bit(R5_UPTODATE, &pdev->flags) ||
4798                                  test_bit(R5_Discard, &pdev->flags))))) &&
4799             (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4800                              && !test_bit(R5_LOCKED, &qdev->flags)
4801                              && (test_bit(R5_UPTODATE, &qdev->flags) ||
4802                                  test_bit(R5_Discard, &qdev->flags))))))
4803                 handle_stripe_clean_event(conf, sh, disks);
4804
4805         if (s.just_cached)
4806                 r5c_handle_cached_data_endio(conf, sh, disks);
4807         log_stripe_write_finished(sh);
4808
4809         /* Now we might consider reading some blocks, either to check/generate
4810          * parity, or to satisfy requests
4811          * or to load a block that is being partially written.
4812          */
4813         if (s.to_read || s.non_overwrite
4814             || (conf->level == 6 && s.to_write && s.failed)
4815             || (s.syncing && (s.uptodate + s.compute < disks))
4816             || s.replacing
4817             || s.expanding)
4818                 handle_stripe_fill(sh, &s, disks);
4819
4820         /*
4821          * When the stripe finishes full journal write cycle (write to journal
4822          * and raid disk), this is the clean up procedure so it is ready for
4823          * next operation.
4824          */
4825         r5c_finish_stripe_write_out(conf, sh, &s);
4826
4827         /*
4828          * Now to consider new write requests, cache write back and what else,
4829          * if anything should be read.  We do not handle new writes when:
4830          * 1/ A 'write' operation (copy+xor) is already in flight.
4831          * 2/ A 'check' operation is in flight, as it may clobber the parity
4832          *    block.
4833          * 3/ A r5c cache log write is in flight.
4834          */
4835
4836         if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
4837                 if (!r5c_is_writeback(conf->log)) {
4838                         if (s.to_write)
4839                                 handle_stripe_dirtying(conf, sh, &s, disks);
4840                 } else { /* write back cache */
4841                         int ret = 0;
4842
4843                         /* First, try handle writes in caching phase */
4844                         if (s.to_write)
4845                                 ret = r5c_try_caching_write(conf, sh, &s,
4846                                                             disks);
4847                         /*
4848                          * If caching phase failed: ret == -EAGAIN
4849                          *    OR
4850                          * stripe under reclaim: !caching && injournal
4851                          *
4852                          * fall back to handle_stripe_dirtying()
4853                          */
4854                         if (ret == -EAGAIN ||
4855                             /* stripe under reclaim: !caching && injournal */
4856                             (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
4857                              s.injournal > 0)) {
4858                                 ret = handle_stripe_dirtying(conf, sh, &s,
4859                                                              disks);
4860                                 if (ret == -EAGAIN)
4861                                         goto finish;
4862                         }
4863                 }
4864         }
4865
4866         /* maybe we need to check and possibly fix the parity for this stripe
4867          * Any reads will already have been scheduled, so we just see if enough
4868          * data is available.  The parity check is held off while parity
4869          * dependent operations are in flight.
4870          */
4871         if (sh->check_state ||
4872             (s.syncing && s.locked == 0 &&
4873              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4874              !test_bit(STRIPE_INSYNC, &sh->state))) {
4875                 if (conf->level == 6)
4876                         handle_parity_checks6(conf, sh, &s, disks);
4877                 else
4878                         handle_parity_checks5(conf, sh, &s, disks);
4879         }
4880
4881         if ((s.replacing || s.syncing) && s.locked == 0
4882             && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4883             && !test_bit(STRIPE_REPLACED, &sh->state)) {
4884                 /* Write out to replacement devices where possible */
4885                 for (i = 0; i < conf->raid_disks; i++)
4886                         if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4887                                 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4888                                 set_bit(R5_WantReplace, &sh->dev[i].flags);
4889                                 set_bit(R5_LOCKED, &sh->dev[i].flags);
4890                                 s.locked++;
4891                         }
4892                 if (s.replacing)
4893                         set_bit(STRIPE_INSYNC, &sh->state);
4894                 set_bit(STRIPE_REPLACED, &sh->state);
4895         }
4896         if ((s.syncing || s.replacing) && s.locked == 0 &&
4897             !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4898             test_bit(STRIPE_INSYNC, &sh->state)) {
4899                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4900                 clear_bit(STRIPE_SYNCING, &sh->state);
4901                 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4902                         wake_up(&conf->wait_for_overlap);
4903         }
4904
4905         /* If the failed drives are just a ReadError, then we might need
4906          * to progress the repair/check process
4907          */
4908         if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4909                 for (i = 0; i < s.failed; i++) {
4910                         struct r5dev *dev = &sh->dev[s.failed_num[i]];
4911                         if (test_bit(R5_ReadError, &dev->flags)
4912                             && !test_bit(R5_LOCKED, &dev->flags)
4913                             && test_bit(R5_UPTODATE, &dev->flags)
4914                                 ) {
4915                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
4916                                         set_bit(R5_Wantwrite, &dev->flags);
4917                                         set_bit(R5_ReWrite, &dev->flags);
4918                                         set_bit(R5_LOCKED, &dev->flags);
4919                                         s.locked++;
4920                                 } else {
4921                                         /* let's read it back */
4922                                         set_bit(R5_Wantread, &dev->flags);
4923                                         set_bit(R5_LOCKED, &dev->flags);
4924                                         s.locked++;
4925                                 }
4926                         }
4927                 }
4928
4929         /* Finish reconstruct operations initiated by the expansion process */
4930         if (sh->reconstruct_state == reconstruct_state_result) {
4931                 struct stripe_head *sh_src
4932                         = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4933                 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4934                         /* sh cannot be written until sh_src has been read.
4935                          * so arrange for sh to be delayed a little
4936                          */
4937                         set_bit(STRIPE_DELAYED, &sh->state);
4938                         set_bit(STRIPE_HANDLE, &sh->state);
4939                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4940                                               &sh_src->state))
4941                                 atomic_inc(&conf->preread_active_stripes);
4942                         raid5_release_stripe(sh_src);
4943                         goto finish;
4944                 }
4945                 if (sh_src)
4946                         raid5_release_stripe(sh_src);
4947
4948                 sh->reconstruct_state = reconstruct_state_idle;
4949                 clear_bit(STRIPE_EXPANDING, &sh->state);
4950                 for (i = conf->raid_disks; i--; ) {
4951                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
4952                         set_bit(R5_LOCKED, &sh->dev[i].flags);
4953                         s.locked++;
4954                 }
4955         }
4956
4957         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4958             !sh->reconstruct_state) {
4959                 /* Need to write out all blocks after computing parity */
4960                 sh->disks = conf->raid_disks;
4961                 stripe_set_idx(sh->sector, conf, 0, sh);
4962                 schedule_reconstruction(sh, &s, 1, 1);
4963         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4964                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4965                 atomic_dec(&conf->reshape_stripes);
4966                 wake_up(&conf->wait_for_overlap);
4967                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4968         }
4969
4970         if (s.expanding && s.locked == 0 &&
4971             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4972                 handle_stripe_expansion(conf, sh);
4973
4974 finish:
4975         /* wait for this device to become unblocked */
4976         if (unlikely(s.blocked_rdev)) {
4977                 if (conf->mddev->external)
4978                         md_wait_for_blocked_rdev(s.blocked_rdev,
4979                                                  conf->mddev);
4980                 else
4981                         /* Internal metadata will immediately
4982                          * be written by raid5d, so we don't
4983                          * need to wait here.
4984                          */
4985                         rdev_dec_pending(s.blocked_rdev,
4986                                          conf->mddev);
4987         }
4988
4989         if (s.handle_bad_blocks)
4990                 for (i = disks; i--; ) {
4991                         struct md_rdev *rdev;
4992                         struct r5dev *dev = &sh->dev[i];
4993                         if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4994                                 /* We own a safe reference to the rdev */
4995                                 rdev = conf->disks[i].rdev;
4996                                 if (!rdev_set_badblocks(rdev, sh->sector,
4997                                                         STRIPE_SECTORS, 0))
4998                                         md_error(conf->mddev, rdev);
4999                                 rdev_dec_pending(rdev, conf->mddev);
5000                         }
5001                         if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
5002                                 rdev = conf->disks[i].rdev;
5003                                 rdev_clear_badblocks(rdev, sh->sector,
5004                                                      STRIPE_SECTORS, 0);
5005                                 rdev_dec_pending(rdev, conf->mddev);
5006                         }
5007                         if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
5008                                 rdev = conf->disks[i].replacement;
5009                                 if (!rdev)
5010                                         /* rdev have been moved down */
5011                                         rdev = conf->disks[i].rdev;
5012                                 rdev_clear_badblocks(rdev, sh->sector,
5013                                                      STRIPE_SECTORS, 0);
5014                                 rdev_dec_pending(rdev, conf->mddev);
5015                         }
5016                 }
5017
5018         if (s.ops_request)
5019                 raid_run_ops(sh, s.ops_request);
5020
5021         ops_run_io(sh, &s);
5022
5023         if (s.dec_preread_active) {
5024                 /* We delay this until after ops_run_io so that if make_request
5025                  * is waiting on a flush, it won't continue until the writes
5026                  * have actually been submitted.
5027                  */
5028                 atomic_dec(&conf->preread_active_stripes);
5029                 if (atomic_read(&conf->preread_active_stripes) <
5030                     IO_THRESHOLD)
5031                         md_wakeup_thread(conf->mddev->thread);
5032         }
5033
5034         clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
5035 }
5036
5037 static void raid5_activate_delayed(struct r5conf *conf)
5038 {
5039         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
5040                 while (!list_empty(&conf->delayed_list)) {
5041                         struct list_head *l = conf->delayed_list.next;
5042                         struct stripe_head *sh;
5043                         sh = list_entry(l, struct stripe_head, lru);
5044                         list_del_init(l);
5045                         clear_bit(STRIPE_DELAYED, &sh->state);
5046                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5047                                 atomic_inc(&conf->preread_active_stripes);
5048                         list_add_tail(&sh->lru, &conf->hold_list);
5049                         raid5_wakeup_stripe_thread(sh);
5050                 }
5051         }
5052 }
5053
5054 static void activate_bit_delay(struct r5conf *conf,
5055         struct list_head *temp_inactive_list)
5056 {
5057         /* device_lock is held */
5058         struct list_head head;
5059         list_add(&head, &conf->bitmap_list);
5060         list_del_init(&conf->bitmap_list);
5061         while (!list_empty(&head)) {
5062                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
5063                 int hash;
5064                 list_del_init(&sh->lru);
5065                 atomic_inc(&sh->count);
5066                 hash = sh->hash_lock_index;
5067                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
5068         }
5069 }
5070
5071 static int raid5_congested(struct mddev *mddev, int bits)
5072 {
5073         struct r5conf *conf = mddev->private;
5074
5075         /* No difference between reads and writes.  Just check
5076          * how busy the stripe_cache is
5077          */
5078
5079         if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
5080                 return 1;
5081
5082         /* Also checks whether there is pressure on r5cache log space */
5083         if (test_bit(R5C_LOG_TIGHT, &conf->cache_state))
5084                 return 1;
5085         if (conf->quiesce)
5086                 return 1;
5087         if (atomic_read(&conf->empty_inactive_list_nr))
5088                 return 1;
5089
5090         return 0;
5091 }
5092
5093 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
5094 {
5095         struct r5conf *conf = mddev->private;
5096         sector_t sector = bio->bi_iter.bi_sector;
5097         unsigned int chunk_sectors;
5098         unsigned int bio_sectors = bio_sectors(bio);
5099
5100         WARN_ON_ONCE(bio->bi_partno);
5101
5102         chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
5103         return  chunk_sectors >=
5104                 ((sector & (chunk_sectors - 1)) + bio_sectors);
5105 }
5106
5107 /*
5108  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
5109  *  later sampled by raid5d.
5110  */
5111 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
5112 {
5113         unsigned long flags;
5114
5115         spin_lock_irqsave(&conf->device_lock, flags);
5116
5117         bi->bi_next = conf->retry_read_aligned_list;
5118         conf->retry_read_aligned_list = bi;
5119
5120         spin_unlock_irqrestore(&conf->device_lock, flags);
5121         md_wakeup_thread(conf->mddev->thread);
5122 }
5123
5124 static struct bio *remove_bio_from_retry(struct r5conf *conf,
5125                                          unsigned int *offset)
5126 {
5127         struct bio *bi;
5128
5129         bi = conf->retry_read_aligned;
5130         if (bi) {
5131                 *offset = conf->retry_read_offset;
5132                 conf->retry_read_aligned = NULL;
5133                 return bi;
5134         }
5135         bi = conf->retry_read_aligned_list;
5136         if(bi) {
5137                 conf->retry_read_aligned_list = bi->bi_next;
5138                 bi->bi_next = NULL;
5139                 *offset = 0;
5140         }
5141
5142         return bi;
5143 }
5144
5145 /*
5146  *  The "raid5_align_endio" should check if the read succeeded and if it
5147  *  did, call bio_endio on the original bio (having bio_put the new bio
5148  *  first).
5149  *  If the read failed..
5150  */
5151 static void raid5_align_endio(struct bio *bi)
5152 {
5153         struct bio* raid_bi  = bi->bi_private;
5154         struct mddev *mddev;
5155         struct r5conf *conf;
5156         struct md_rdev *rdev;
5157         blk_status_t error = bi->bi_status;
5158
5159         bio_put(bi);
5160
5161         rdev = (void*)raid_bi->bi_next;
5162         raid_bi->bi_next = NULL;
5163         mddev = rdev->mddev;
5164         conf = mddev->private;
5165
5166         rdev_dec_pending(rdev, conf->mddev);
5167
5168         if (!error) {
5169                 bio_endio(raid_bi);
5170                 if (atomic_dec_and_test(&conf->active_aligned_reads))
5171                         wake_up(&conf->wait_for_quiescent);
5172                 return;
5173         }
5174
5175         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5176
5177         add_bio_to_retry(raid_bi, conf);
5178 }
5179
5180 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
5181 {
5182         struct r5conf *conf = mddev->private;
5183         int dd_idx;
5184         struct bio* align_bi;
5185         struct md_rdev *rdev;
5186         sector_t end_sector;
5187
5188         if (!in_chunk_boundary(mddev, raid_bio)) {
5189                 pr_debug("%s: non aligned\n", __func__);
5190                 return 0;
5191         }
5192         /*
5193          * use bio_clone_fast to make a copy of the bio
5194          */
5195         align_bi = bio_clone_fast(raid_bio, GFP_NOIO, mddev->bio_set);
5196         if (!align_bi)
5197                 return 0;
5198         /*
5199          *   set bi_end_io to a new function, and set bi_private to the
5200          *     original bio.
5201          */
5202         align_bi->bi_end_io  = raid5_align_endio;
5203         align_bi->bi_private = raid_bio;
5204         /*
5205          *      compute position
5206          */
5207         align_bi->bi_iter.bi_sector =
5208                 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
5209                                      0, &dd_idx, NULL);
5210
5211         end_sector = bio_end_sector(align_bi);
5212         rcu_read_lock();
5213         rdev = rcu_dereference(conf->disks[dd_idx].replacement);
5214         if (!rdev || test_bit(Faulty, &rdev->flags) ||
5215             rdev->recovery_offset < end_sector) {
5216                 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
5217                 if (rdev &&
5218                     (test_bit(Faulty, &rdev->flags) ||
5219                     !(test_bit(In_sync, &rdev->flags) ||
5220                       rdev->recovery_offset >= end_sector)))
5221                         rdev = NULL;
5222         }
5223
5224         if (r5c_big_stripe_cached(conf, align_bi->bi_iter.bi_sector)) {
5225                 rcu_read_unlock();
5226                 bio_put(align_bi);
5227                 return 0;
5228         }
5229
5230         if (rdev) {
5231                 sector_t first_bad;
5232                 int bad_sectors;
5233
5234                 atomic_inc(&rdev->nr_pending);
5235                 rcu_read_unlock();
5236                 raid_bio->bi_next = (void*)rdev;
5237                 bio_set_dev(align_bi, rdev->bdev);
5238                 bio_clear_flag(align_bi, BIO_SEG_VALID);
5239
5240                 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
5241                                 bio_sectors(align_bi),
5242                                 &first_bad, &bad_sectors)) {
5243                         bio_put(align_bi);
5244                         rdev_dec_pending(rdev, mddev);
5245                         return 0;
5246                 }
5247
5248                 /* No reshape active, so we can trust rdev->data_offset */
5249                 align_bi->bi_iter.bi_sector += rdev->data_offset;
5250
5251                 spin_lock_irq(&conf->device_lock);
5252                 wait_event_lock_irq(conf->wait_for_quiescent,
5253                                     conf->quiesce == 0,
5254                                     conf->device_lock);
5255                 atomic_inc(&conf->active_aligned_reads);
5256                 spin_unlock_irq(&conf->device_lock);
5257
5258                 if (mddev->gendisk)
5259                         trace_block_bio_remap(align_bi->bi_disk->queue,
5260                                               align_bi, disk_devt(mddev->gendisk),
5261                                               raid_bio->bi_iter.bi_sector);
5262                 generic_make_request(align_bi);
5263                 return 1;
5264         } else {
5265                 rcu_read_unlock();
5266                 bio_put(align_bi);
5267                 return 0;
5268         }
5269 }
5270
5271 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
5272 {
5273         struct bio *split;
5274         sector_t sector = raid_bio->bi_iter.bi_sector;
5275         unsigned chunk_sects = mddev->chunk_sectors;
5276         unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
5277
5278         if (sectors < bio_sectors(raid_bio)) {
5279                 struct r5conf *conf = mddev->private;
5280                 split = bio_split(raid_bio, sectors, GFP_NOIO, conf->bio_split);
5281                 bio_chain(split, raid_bio);
5282                 generic_make_request(raid_bio);
5283                 raid_bio = split;
5284         }
5285
5286         if (!raid5_read_one_chunk(mddev, raid_bio))
5287                 return raid_bio;
5288
5289         return NULL;
5290 }
5291
5292 /* __get_priority_stripe - get the next stripe to process
5293  *
5294  * Full stripe writes are allowed to pass preread active stripes up until
5295  * the bypass_threshold is exceeded.  In general the bypass_count
5296  * increments when the handle_list is handled before the hold_list; however, it
5297  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5298  * stripe with in flight i/o.  The bypass_count will be reset when the
5299  * head of the hold_list has changed, i.e. the head was promoted to the
5300  * handle_list.
5301  */
5302 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
5303 {
5304         struct stripe_head *sh, *tmp;
5305         struct list_head *handle_list = NULL;
5306         struct r5worker_group *wg;
5307         bool second_try = !r5c_is_writeback(conf->log) &&
5308                 !r5l_log_disk_error(conf);
5309         bool try_loprio = test_bit(R5C_LOG_TIGHT, &conf->cache_state) ||
5310                 r5l_log_disk_error(conf);
5311
5312 again:
5313         wg = NULL;
5314         sh = NULL;
5315         if (conf->worker_cnt_per_group == 0) {
5316                 handle_list = try_loprio ? &conf->loprio_list :
5317                                         &conf->handle_list;
5318         } else if (group != ANY_GROUP) {
5319                 handle_list = try_loprio ? &conf->worker_groups[group].loprio_list :
5320                                 &conf->worker_groups[group].handle_list;
5321                 wg = &conf->worker_groups[group];
5322         } else {
5323                 int i;
5324                 for (i = 0; i < conf->group_cnt; i++) {
5325                         handle_list = try_loprio ? &conf->worker_groups[i].loprio_list :
5326                                 &conf->worker_groups[i].handle_list;
5327                         wg = &conf->worker_groups[i];
5328                         if (!list_empty(handle_list))
5329                                 break;
5330                 }
5331         }
5332
5333         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5334                   __func__,
5335                   list_empty(handle_list) ? "empty" : "busy",
5336                   list_empty(&conf->hold_list) ? "empty" : "busy",
5337                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
5338
5339         if (!list_empty(handle_list)) {
5340                 sh = list_entry(handle_list->next, typeof(*sh), lru);
5341
5342                 if (list_empty(&conf->hold_list))
5343                         conf->bypass_count = 0;
5344                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
5345                         if (conf->hold_list.next == conf->last_hold)
5346                                 conf->bypass_count++;
5347                         else {
5348                                 conf->last_hold = conf->hold_list.next;
5349                                 conf->bypass_count -= conf->bypass_threshold;
5350                                 if (conf->bypass_count < 0)
5351                                         conf->bypass_count = 0;
5352                         }
5353                 }
5354         } else if (!list_empty(&conf->hold_list) &&
5355                    ((conf->bypass_threshold &&
5356                      conf->bypass_count > conf->bypass_threshold) ||
5357                     atomic_read(&conf->pending_full_writes) == 0)) {
5358
5359                 list_for_each_entry(tmp, &conf->hold_list,  lru) {
5360                         if (conf->worker_cnt_per_group == 0 ||
5361                             group == ANY_GROUP ||
5362                             !cpu_online(tmp->cpu) ||
5363                             cpu_to_group(tmp->cpu) == group) {
5364                                 sh = tmp;
5365                                 break;
5366                         }
5367                 }
5368
5369                 if (sh) {
5370                         conf->bypass_count -= conf->bypass_threshold;
5371                         if (conf->bypass_count < 0)
5372                                 conf->bypass_count = 0;
5373                 }
5374                 wg = NULL;
5375         }
5376
5377         if (!sh) {
5378                 if (second_try)
5379                         return NULL;
5380                 second_try = true;
5381                 try_loprio = !try_loprio;
5382                 goto again;
5383         }
5384
5385         if (wg) {
5386                 wg->stripes_cnt--;
5387                 sh->group = NULL;
5388         }
5389         list_del_init(&sh->lru);
5390         BUG_ON(atomic_inc_return(&sh->count) != 1);
5391         return sh;
5392 }
5393
5394 struct raid5_plug_cb {
5395         struct blk_plug_cb      cb;
5396         struct list_head        list;
5397         struct list_head        temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5398 };
5399
5400 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5401 {
5402         struct raid5_plug_cb *cb = container_of(
5403                 blk_cb, struct raid5_plug_cb, cb);
5404         struct stripe_head *sh;
5405         struct mddev *mddev = cb->cb.data;
5406         struct r5conf *conf = mddev->private;
5407         int cnt = 0;
5408         int hash;
5409
5410         if (cb->list.next && !list_empty(&cb->list)) {
5411                 spin_lock_irq(&conf->device_lock);
5412                 while (!list_empty(&cb->list)) {
5413                         sh = list_first_entry(&cb->list, struct stripe_head, lru);
5414                         list_del_init(&sh->lru);
5415                         /*
5416                          * avoid race release_stripe_plug() sees
5417                          * STRIPE_ON_UNPLUG_LIST clear but the stripe
5418                          * is still in our list
5419                          */
5420                         smp_mb__before_atomic();
5421                         clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5422                         /*
5423                          * STRIPE_ON_RELEASE_LIST could be set here. In that
5424                          * case, the count is always > 1 here
5425                          */
5426                         hash = sh->hash_lock_index;
5427                         __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5428                         cnt++;
5429                 }
5430                 spin_unlock_irq(&conf->device_lock);
5431         }
5432         release_inactive_stripe_list(conf, cb->temp_inactive_list,
5433                                      NR_STRIPE_HASH_LOCKS);
5434         if (mddev->queue)
5435                 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5436         kfree(cb);
5437 }
5438
5439 static void release_stripe_plug(struct mddev *mddev,
5440                                 struct stripe_head *sh)
5441 {
5442         struct blk_plug_cb *blk_cb = blk_check_plugged(
5443                 raid5_unplug, mddev,
5444                 sizeof(struct raid5_plug_cb));
5445         struct raid5_plug_cb *cb;
5446
5447         if (!blk_cb) {
5448                 raid5_release_stripe(sh);
5449                 return;
5450         }
5451
5452         cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5453
5454         if (cb->list.next == NULL) {
5455                 int i;
5456                 INIT_LIST_HEAD(&cb->list);
5457                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5458                         INIT_LIST_HEAD(cb->temp_inactive_list + i);
5459         }
5460
5461         if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5462                 list_add_tail(&sh->lru, &cb->list);
5463         else
5464                 raid5_release_stripe(sh);
5465 }
5466
5467 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5468 {
5469         struct r5conf *conf = mddev->private;
5470         sector_t logical_sector, last_sector;
5471         struct stripe_head *sh;
5472         int stripe_sectors;
5473
5474         if (mddev->reshape_position != MaxSector)
5475                 /* Skip discard while reshape is happening */
5476                 return;
5477
5478         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5479         last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5480
5481         bi->bi_next = NULL;
5482
5483         stripe_sectors = conf->chunk_sectors *
5484                 (conf->raid_disks - conf->max_degraded);
5485         logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5486                                                stripe_sectors);
5487         sector_div(last_sector, stripe_sectors);
5488
5489         logical_sector *= conf->chunk_sectors;
5490         last_sector *= conf->chunk_sectors;
5491
5492         for (; logical_sector < last_sector;
5493              logical_sector += STRIPE_SECTORS) {
5494                 DEFINE_WAIT(w);
5495                 int d;
5496         again:
5497                 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5498                 prepare_to_wait(&conf->wait_for_overlap, &w,
5499                                 TASK_UNINTERRUPTIBLE);
5500                 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5501                 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5502                         raid5_release_stripe(sh);
5503                         schedule();
5504                         goto again;
5505                 }
5506                 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5507                 spin_lock_irq(&sh->stripe_lock);
5508                 for (d = 0; d < conf->raid_disks; d++) {
5509                         if (d == sh->pd_idx || d == sh->qd_idx)
5510                                 continue;
5511                         if (sh->dev[d].towrite || sh->dev[d].toread) {
5512                                 set_bit(R5_Overlap, &sh->dev[d].flags);
5513                                 spin_unlock_irq(&sh->stripe_lock);
5514                                 raid5_release_stripe(sh);
5515                                 schedule();
5516                                 goto again;
5517                         }
5518                 }
5519                 set_bit(STRIPE_DISCARD, &sh->state);
5520                 finish_wait(&conf->wait_for_overlap, &w);
5521                 sh->overwrite_disks = 0;
5522                 for (d = 0; d < conf->raid_disks; d++) {
5523                         if (d == sh->pd_idx || d == sh->qd_idx)
5524                                 continue;
5525                         sh->dev[d].towrite = bi;
5526                         set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5527                         bio_inc_remaining(bi);
5528                         md_write_inc(mddev, bi);
5529                         sh->overwrite_disks++;
5530                 }
5531                 spin_unlock_irq(&sh->stripe_lock);
5532                 if (conf->mddev->bitmap) {
5533                         for (d = 0;
5534                              d < conf->raid_disks - conf->max_degraded;
5535                              d++)
5536                                 bitmap_startwrite(mddev->bitmap,
5537                                                   sh->sector,
5538                                                   STRIPE_SECTORS,
5539                                                   0);
5540                         sh->bm_seq = conf->seq_flush + 1;
5541                         set_bit(STRIPE_BIT_DELAY, &sh->state);
5542                 }
5543
5544                 set_bit(STRIPE_HANDLE, &sh->state);
5545                 clear_bit(STRIPE_DELAYED, &sh->state);
5546                 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5547                         atomic_inc(&conf->preread_active_stripes);
5548                 release_stripe_plug(mddev, sh);
5549         }
5550
5551         bio_endio(bi);
5552 }
5553
5554 static bool raid5_make_request(struct mddev *mddev, struct bio * bi)
5555 {
5556         struct r5conf *conf = mddev->private;
5557         int dd_idx;
5558         sector_t new_sector;
5559         sector_t logical_sector, last_sector;
5560         struct stripe_head *sh;
5561         const int rw = bio_data_dir(bi);
5562         DEFINE_WAIT(w);
5563         bool do_prepare;
5564         bool do_flush = false;
5565
5566         if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
5567                 int ret = r5l_handle_flush_request(conf->log, bi);
5568
5569                 if (ret == 0)
5570                         return true;
5571                 if (ret == -ENODEV) {
5572                         md_flush_request(mddev, bi);
5573                         return true;
5574                 }
5575                 /* ret == -EAGAIN, fallback */
5576                 /*
5577                  * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5578                  * we need to flush journal device
5579                  */
5580                 do_flush = bi->bi_opf & REQ_PREFLUSH;
5581         }
5582
5583         if (!md_write_start(mddev, bi))
5584                 return false;
5585         /*
5586          * If array is degraded, better not do chunk aligned read because
5587          * later we might have to read it again in order to reconstruct
5588          * data on failed drives.
5589          */
5590         if (rw == READ && mddev->degraded == 0 &&
5591             mddev->reshape_position == MaxSector) {
5592                 bi = chunk_aligned_read(mddev, bi);
5593                 if (!bi)
5594                         return true;
5595         }
5596
5597         if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
5598                 make_discard_request(mddev, bi);
5599                 md_write_end(mddev);
5600                 return true;
5601         }
5602
5603         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5604         last_sector = bio_end_sector(bi);
5605         bi->bi_next = NULL;
5606
5607         prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5608         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5609                 int previous;
5610                 int seq;
5611
5612                 do_prepare = false;
5613         retry:
5614                 seq = read_seqcount_begin(&conf->gen_lock);
5615                 previous = 0;
5616                 if (do_prepare)
5617                         prepare_to_wait(&conf->wait_for_overlap, &w,
5618                                 TASK_UNINTERRUPTIBLE);
5619                 if (unlikely(conf->reshape_progress != MaxSector)) {
5620                         /* spinlock is needed as reshape_progress may be
5621                          * 64bit on a 32bit platform, and so it might be
5622                          * possible to see a half-updated value
5623                          * Of course reshape_progress could change after
5624                          * the lock is dropped, so once we get a reference
5625                          * to the stripe that we think it is, we will have
5626                          * to check again.
5627                          */
5628                         spin_lock_irq(&conf->device_lock);
5629                         if (mddev->reshape_backwards
5630                             ? logical_sector < conf->reshape_progress
5631                             : logical_sector >= conf->reshape_progress) {
5632                                 previous = 1;
5633                         } else {
5634                                 if (mddev->reshape_backwards
5635                                     ? logical_sector < conf->reshape_safe
5636                                     : logical_sector >= conf->reshape_safe) {
5637                                         spin_unlock_irq(&conf->device_lock);
5638                                         schedule();
5639                                         do_prepare = true;
5640                                         goto retry;
5641                                 }
5642                         }
5643                         spin_unlock_irq(&conf->device_lock);
5644                 }
5645
5646                 new_sector = raid5_compute_sector(conf, logical_sector,
5647                                                   previous,
5648                                                   &dd_idx, NULL);
5649                 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5650                         (unsigned long long)new_sector,
5651                         (unsigned long long)logical_sector);
5652
5653                 sh = raid5_get_active_stripe(conf, new_sector, previous,
5654                                        (bi->bi_opf & REQ_RAHEAD), 0);
5655                 if (sh) {
5656                         if (unlikely(previous)) {
5657                                 /* expansion might have moved on while waiting for a
5658                                  * stripe, so we must do the range check again.
5659                                  * Expansion could still move past after this
5660                                  * test, but as we are holding a reference to
5661                                  * 'sh', we know that if that happens,
5662                                  *  STRIPE_EXPANDING will get set and the expansion
5663                                  * won't proceed until we finish with the stripe.
5664                                  */
5665                                 int must_retry = 0;
5666                                 spin_lock_irq(&conf->device_lock);
5667                                 if (mddev->reshape_backwards
5668                                     ? logical_sector >= conf->reshape_progress
5669                                     : logical_sector < conf->reshape_progress)
5670                                         /* mismatch, need to try again */
5671                                         must_retry = 1;
5672                                 spin_unlock_irq(&conf->device_lock);
5673                                 if (must_retry) {
5674                                         raid5_release_stripe(sh);
5675                                         schedule();
5676                                         do_prepare = true;
5677                                         goto retry;
5678                                 }
5679                         }
5680                         if (read_seqcount_retry(&conf->gen_lock, seq)) {
5681                                 /* Might have got the wrong stripe_head
5682                                  * by accident
5683                                  */
5684                                 raid5_release_stripe(sh);
5685                                 goto retry;
5686                         }
5687
5688                         if (rw == WRITE &&
5689                             logical_sector >= mddev->suspend_lo &&
5690                             logical_sector < mddev->suspend_hi) {
5691                                 raid5_release_stripe(sh);
5692                                 /* As the suspend_* range is controlled by
5693                                  * userspace, we want an interruptible
5694                                  * wait.
5695                                  */
5696                                 prepare_to_wait(&conf->wait_for_overlap,
5697                                                 &w, TASK_INTERRUPTIBLE);
5698                                 if (logical_sector >= mddev->suspend_lo &&
5699                                     logical_sector < mddev->suspend_hi) {
5700                                         sigset_t full, old;
5701                                         sigfillset(&full);
5702                                         sigprocmask(SIG_BLOCK, &full, &old);
5703                                         schedule();
5704                                         sigprocmask(SIG_SETMASK, &old, NULL);
5705                                         do_prepare = true;
5706                                 }
5707                                 goto retry;
5708                         }
5709
5710                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5711                             !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5712                                 /* Stripe is busy expanding or
5713                                  * add failed due to overlap.  Flush everything
5714                                  * and wait a while
5715                                  */
5716                                 md_wakeup_thread(mddev->thread);
5717                                 raid5_release_stripe(sh);
5718                                 schedule();
5719                                 do_prepare = true;
5720                                 goto retry;
5721                         }
5722                         if (do_flush) {
5723                                 set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
5724                                 /* we only need flush for one stripe */
5725                                 do_flush = false;
5726                         }
5727
5728                         set_bit(STRIPE_HANDLE, &sh->state);
5729                         clear_bit(STRIPE_DELAYED, &sh->state);
5730                         if ((!sh->batch_head || sh == sh->batch_head) &&
5731                             (bi->bi_opf & REQ_SYNC) &&
5732                             !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5733                                 atomic_inc(&conf->preread_active_stripes);
5734                         release_stripe_plug(mddev, sh);
5735                 } else {
5736                         /* cannot get stripe for read-ahead, just give-up */
5737                         bi->bi_status = BLK_STS_IOERR;
5738                         break;
5739                 }
5740         }
5741         finish_wait(&conf->wait_for_overlap, &w);
5742
5743         if (rw == WRITE)
5744                 md_write_end(mddev);
5745         bio_endio(bi);
5746         return true;
5747 }
5748
5749 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5750
5751 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5752 {
5753         /* reshaping is quite different to recovery/resync so it is
5754          * handled quite separately ... here.
5755          *
5756          * On each call to sync_request, we gather one chunk worth of
5757          * destination stripes and flag them as expanding.
5758          * Then we find all the source stripes and request reads.
5759          * As the reads complete, handle_stripe will copy the data
5760          * into the destination stripe and release that stripe.
5761          */
5762         struct r5conf *conf = mddev->private;
5763         struct stripe_head *sh;
5764         sector_t first_sector, last_sector;
5765         int raid_disks = conf->previous_raid_disks;
5766         int data_disks = raid_disks - conf->max_degraded;
5767         int new_data_disks = conf->raid_disks - conf->max_degraded;
5768         int i;
5769         int dd_idx;
5770         sector_t writepos, readpos, safepos;
5771         sector_t stripe_addr;
5772         int reshape_sectors;
5773         struct list_head stripes;
5774         sector_t retn;
5775
5776         if (sector_nr == 0) {
5777                 /* If restarting in the middle, skip the initial sectors */
5778                 if (mddev->reshape_backwards &&
5779                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5780                         sector_nr = raid5_size(mddev, 0, 0)
5781                                 - conf->reshape_progress;
5782                 } else if (mddev->reshape_backwards &&
5783                            conf->reshape_progress == MaxSector) {
5784                         /* shouldn't happen, but just in case, finish up.*/
5785                         sector_nr = MaxSector;
5786                 } else if (!mddev->reshape_backwards &&
5787                            conf->reshape_progress > 0)
5788                         sector_nr = conf->reshape_progress;
5789                 sector_div(sector_nr, new_data_disks);
5790                 if (sector_nr) {
5791                         mddev->curr_resync_completed = sector_nr;
5792                         sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5793                         *skipped = 1;
5794                         retn = sector_nr;
5795                         goto finish;
5796                 }
5797         }
5798
5799         /* We need to process a full chunk at a time.
5800          * If old and new chunk sizes differ, we need to process the
5801          * largest of these
5802          */
5803
5804         reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5805
5806         /* We update the metadata at least every 10 seconds, or when
5807          * the data about to be copied would over-write the source of
5808          * the data at the front of the range.  i.e. one new_stripe
5809          * along from reshape_progress new_maps to after where
5810          * reshape_safe old_maps to
5811          */
5812         writepos = conf->reshape_progress;
5813         sector_div(writepos, new_data_disks);
5814         readpos = conf->reshape_progress;
5815         sector_div(readpos, data_disks);
5816         safepos = conf->reshape_safe;
5817         sector_div(safepos, data_disks);
5818         if (mddev->reshape_backwards) {
5819                 BUG_ON(writepos < reshape_sectors);
5820                 writepos -= reshape_sectors;
5821                 readpos += reshape_sectors;
5822                 safepos += reshape_sectors;
5823         } else {
5824                 writepos += reshape_sectors;
5825                 /* readpos and safepos are worst-case calculations.
5826                  * A negative number is overly pessimistic, and causes
5827                  * obvious problems for unsigned storage.  So clip to 0.
5828                  */
5829                 readpos -= min_t(sector_t, reshape_sectors, readpos);
5830                 safepos -= min_t(sector_t, reshape_sectors, safepos);
5831         }
5832
5833         /* Having calculated the 'writepos' possibly use it
5834          * to set 'stripe_addr' which is where we will write to.
5835          */
5836         if (mddev->reshape_backwards) {
5837                 BUG_ON(conf->reshape_progress == 0);
5838                 stripe_addr = writepos;
5839                 BUG_ON((mddev->dev_sectors &
5840                         ~((sector_t)reshape_sectors - 1))
5841                        - reshape_sectors - stripe_addr
5842                        != sector_nr);
5843         } else {
5844                 BUG_ON(writepos != sector_nr + reshape_sectors);
5845                 stripe_addr = sector_nr;
5846         }
5847
5848         /* 'writepos' is the most advanced device address we might write.
5849          * 'readpos' is the least advanced device address we might read.
5850          * 'safepos' is the least address recorded in the metadata as having
5851          *     been reshaped.
5852          * If there is a min_offset_diff, these are adjusted either by
5853          * increasing the safepos/readpos if diff is negative, or
5854          * increasing writepos if diff is positive.
5855          * If 'readpos' is then behind 'writepos', there is no way that we can
5856          * ensure safety in the face of a crash - that must be done by userspace
5857          * making a backup of the data.  So in that case there is no particular
5858          * rush to update metadata.
5859          * Otherwise if 'safepos' is behind 'writepos', then we really need to
5860          * update the metadata to advance 'safepos' to match 'readpos' so that
5861          * we can be safe in the event of a crash.
5862          * So we insist on updating metadata if safepos is behind writepos and
5863          * readpos is beyond writepos.
5864          * In any case, update the metadata every 10 seconds.
5865          * Maybe that number should be configurable, but I'm not sure it is
5866          * worth it.... maybe it could be a multiple of safemode_delay???
5867          */
5868         if (conf->min_offset_diff < 0) {
5869                 safepos += -conf->min_offset_diff;
5870                 readpos += -conf->min_offset_diff;
5871         } else
5872                 writepos += conf->min_offset_diff;
5873
5874         if ((mddev->reshape_backwards
5875              ? (safepos > writepos && readpos < writepos)
5876              : (safepos < writepos && readpos > writepos)) ||
5877             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5878                 /* Cannot proceed until we've updated the superblock... */
5879                 wait_event(conf->wait_for_overlap,
5880                            atomic_read(&conf->reshape_stripes)==0
5881                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5882                 if (atomic_read(&conf->reshape_stripes) != 0)
5883                         return 0;
5884                 mddev->reshape_position = conf->reshape_progress;
5885                 mddev->curr_resync_completed = sector_nr;
5886                 conf->reshape_checkpoint = jiffies;
5887                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5888                 md_wakeup_thread(mddev->thread);
5889                 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
5890                            test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5891                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5892                         return 0;
5893                 spin_lock_irq(&conf->device_lock);
5894                 conf->reshape_safe = mddev->reshape_position;
5895                 spin_unlock_irq(&conf->device_lock);
5896                 wake_up(&conf->wait_for_overlap);
5897                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5898         }
5899
5900         INIT_LIST_HEAD(&stripes);
5901         for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5902                 int j;
5903                 int skipped_disk = 0;
5904                 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5905                 set_bit(STRIPE_EXPANDING, &sh->state);
5906                 atomic_inc(&conf->reshape_stripes);
5907                 /* If any of this stripe is beyond the end of the old
5908                  * array, then we need to zero those blocks
5909                  */
5910                 for (j=sh->disks; j--;) {
5911                         sector_t s;
5912                         if (j == sh->pd_idx)
5913                                 continue;
5914                         if (conf->level == 6 &&
5915                             j == sh->qd_idx)
5916                                 continue;
5917                         s = raid5_compute_blocknr(sh, j, 0);
5918                         if (s < raid5_size(mddev, 0, 0)) {
5919                                 skipped_disk = 1;
5920                                 continue;
5921                         }
5922                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5923                         set_bit(R5_Expanded, &sh->dev[j].flags);
5924                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
5925                 }
5926                 if (!skipped_disk) {
5927                         set_bit(STRIPE_EXPAND_READY, &sh->state);
5928                         set_bit(STRIPE_HANDLE, &sh->state);
5929                 }
5930                 list_add(&sh->lru, &stripes);
5931         }
5932         spin_lock_irq(&conf->device_lock);
5933         if (mddev->reshape_backwards)
5934                 conf->reshape_progress -= reshape_sectors * new_data_disks;
5935         else
5936                 conf->reshape_progress += reshape_sectors * new_data_disks;
5937         spin_unlock_irq(&conf->device_lock);
5938         /* Ok, those stripe are ready. We can start scheduling
5939          * reads on the source stripes.
5940          * The source stripes are determined by mapping the first and last
5941          * block on the destination stripes.
5942          */
5943         first_sector =
5944                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5945                                      1, &dd_idx, NULL);
5946         last_sector =
5947                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5948                                             * new_data_disks - 1),
5949                                      1, &dd_idx, NULL);
5950         if (last_sector >= mddev->dev_sectors)
5951                 last_sector = mddev->dev_sectors - 1;
5952         while (first_sector <= last_sector) {
5953                 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5954                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5955                 set_bit(STRIPE_HANDLE, &sh->state);
5956                 raid5_release_stripe(sh);
5957                 first_sector += STRIPE_SECTORS;
5958         }
5959         /* Now that the sources are clearly marked, we can release
5960          * the destination stripes
5961          */
5962         while (!list_empty(&stripes)) {
5963                 sh = list_entry(stripes.next, struct stripe_head, lru);
5964                 list_del_init(&sh->lru);
5965                 raid5_release_stripe(sh);
5966         }
5967         /* If this takes us to the resync_max point where we have to pause,
5968          * then we need to write out the superblock.
5969          */
5970         sector_nr += reshape_sectors;
5971         retn = reshape_sectors;
5972 finish:
5973         if (mddev->curr_resync_completed > mddev->resync_max ||
5974             (sector_nr - mddev->curr_resync_completed) * 2
5975             >= mddev->resync_max - mddev->curr_resync_completed) {
5976                 /* Cannot proceed until we've updated the superblock... */
5977                 wait_event(conf->wait_for_overlap,
5978                            atomic_read(&conf->reshape_stripes) == 0
5979                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5980                 if (atomic_read(&conf->reshape_stripes) != 0)
5981                         goto ret;
5982                 mddev->reshape_position = conf->reshape_progress;
5983                 mddev->curr_resync_completed = sector_nr;
5984                 conf->reshape_checkpoint = jiffies;
5985                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5986                 md_wakeup_thread(mddev->thread);
5987                 wait_event(mddev->sb_wait,
5988                            !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
5989                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5990                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5991                         goto ret;
5992                 spin_lock_irq(&conf->device_lock);
5993                 conf->reshape_safe = mddev->reshape_position;
5994                 spin_unlock_irq(&conf->device_lock);
5995                 wake_up(&conf->wait_for_overlap);
5996                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5997         }
5998 ret:
5999         return retn;
6000 }
6001
6002 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
6003                                           int *skipped)
6004 {
6005         struct r5conf *conf = mddev->private;
6006         struct stripe_head *sh;
6007         sector_t max_sector = mddev->dev_sectors;
6008         sector_t sync_blocks;
6009         int still_degraded = 0;
6010         int i;
6011
6012         if (sector_nr >= max_sector) {
6013                 /* just being told to finish up .. nothing much to do */
6014
6015                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
6016                         end_reshape(conf);
6017                         return 0;
6018                 }
6019
6020                 if (mddev->curr_resync < max_sector) /* aborted */
6021                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
6022                                         &sync_blocks, 1);
6023                 else /* completed sync */
6024                         conf->fullsync = 0;
6025                 bitmap_close_sync(mddev->bitmap);
6026
6027                 return 0;
6028         }
6029
6030         /* Allow raid5_quiesce to complete */
6031         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
6032
6033         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
6034                 return reshape_request(mddev, sector_nr, skipped);
6035
6036         /* No need to check resync_max as we never do more than one
6037          * stripe, and as resync_max will always be on a chunk boundary,
6038          * if the check in md_do_sync didn't fire, there is no chance
6039          * of overstepping resync_max here
6040          */
6041
6042         /* if there is too many failed drives and we are trying
6043          * to resync, then assert that we are finished, because there is
6044          * nothing we can do.
6045          */
6046         if (mddev->degraded >= conf->max_degraded &&
6047             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
6048                 sector_t rv = mddev->dev_sectors - sector_nr;
6049                 *skipped = 1;
6050                 return rv;
6051         }
6052         if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
6053             !conf->fullsync &&
6054             !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
6055             sync_blocks >= STRIPE_SECTORS) {
6056                 /* we can skip this block, and probably more */
6057                 sync_blocks /= STRIPE_SECTORS;
6058                 *skipped = 1;
6059                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
6060         }
6061
6062         bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
6063
6064         sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
6065         if (sh == NULL) {
6066                 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
6067                 /* make sure we don't swamp the stripe cache if someone else
6068                  * is trying to get access
6069                  */
6070                 schedule_timeout_uninterruptible(1);
6071         }
6072         /* Need to check if array will still be degraded after recovery/resync
6073          * Note in case of > 1 drive failures it's possible we're rebuilding
6074          * one drive while leaving another faulty drive in array.
6075          */
6076         rcu_read_lock();
6077         for (i = 0; i < conf->raid_disks; i++) {
6078                 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
6079
6080                 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
6081                         still_degraded = 1;
6082         }
6083         rcu_read_unlock();
6084
6085         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
6086
6087         set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
6088         set_bit(STRIPE_HANDLE, &sh->state);
6089
6090         raid5_release_stripe(sh);
6091
6092         return STRIPE_SECTORS;
6093 }
6094
6095 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio,
6096                                unsigned int offset)
6097 {
6098         /* We may not be able to submit a whole bio at once as there
6099          * may not be enough stripe_heads available.
6100          * We cannot pre-allocate enough stripe_heads as we may need
6101          * more than exist in the cache (if we allow ever large chunks).
6102          * So we do one stripe head at a time and record in
6103          * ->bi_hw_segments how many have been done.
6104          *
6105          * We *know* that this entire raid_bio is in one chunk, so
6106          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6107          */
6108         struct stripe_head *sh;
6109         int dd_idx;
6110         sector_t sector, logical_sector, last_sector;
6111         int scnt = 0;
6112         int handled = 0;
6113
6114         logical_sector = raid_bio->bi_iter.bi_sector &
6115                 ~((sector_t)STRIPE_SECTORS-1);
6116         sector = raid5_compute_sector(conf, logical_sector,
6117                                       0, &dd_idx, NULL);
6118         last_sector = bio_end_sector(raid_bio);
6119
6120         for (; logical_sector < last_sector;
6121              logical_sector += STRIPE_SECTORS,
6122                      sector += STRIPE_SECTORS,
6123                      scnt++) {
6124
6125                 if (scnt < offset)
6126                         /* already done this stripe */
6127                         continue;
6128
6129                 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
6130
6131                 if (!sh) {
6132                         /* failed to get a stripe - must wait */
6133                         conf->retry_read_aligned = raid_bio;
6134                         conf->retry_read_offset = scnt;
6135                         return handled;
6136                 }
6137
6138                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
6139                         raid5_release_stripe(sh);
6140                         conf->retry_read_aligned = raid_bio;
6141                         conf->retry_read_offset = scnt;
6142                         return handled;
6143                 }
6144
6145                 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
6146                 handle_stripe(sh);
6147                 raid5_release_stripe(sh);
6148                 handled++;
6149         }
6150
6151         bio_endio(raid_bio);
6152
6153         if (atomic_dec_and_test(&conf->active_aligned_reads))
6154                 wake_up(&conf->wait_for_quiescent);
6155         return handled;
6156 }
6157
6158 static int handle_active_stripes(struct r5conf *conf, int group,
6159                                  struct r5worker *worker,
6160                                  struct list_head *temp_inactive_list)
6161 {
6162         struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
6163         int i, batch_size = 0, hash;
6164         bool release_inactive = false;
6165
6166         while (batch_size < MAX_STRIPE_BATCH &&
6167                         (sh = __get_priority_stripe(conf, group)) != NULL)
6168                 batch[batch_size++] = sh;
6169
6170         if (batch_size == 0) {
6171                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6172                         if (!list_empty(temp_inactive_list + i))
6173                                 break;
6174                 if (i == NR_STRIPE_HASH_LOCKS) {
6175                         spin_unlock_irq(&conf->device_lock);
6176                         r5l_flush_stripe_to_raid(conf->log);
6177                         spin_lock_irq(&conf->device_lock);
6178                         return batch_size;
6179                 }
6180                 release_inactive = true;
6181         }
6182         spin_unlock_irq(&conf->device_lock);
6183
6184         release_inactive_stripe_list(conf, temp_inactive_list,
6185                                      NR_STRIPE_HASH_LOCKS);
6186
6187         r5l_flush_stripe_to_raid(conf->log);
6188         if (release_inactive) {
6189                 spin_lock_irq(&conf->device_lock);
6190                 return 0;
6191         }
6192
6193         for (i = 0; i < batch_size; i++)
6194                 handle_stripe(batch[i]);
6195         log_write_stripe_run(conf);
6196
6197         cond_resched();
6198
6199         spin_lock_irq(&conf->device_lock);
6200         for (i = 0; i < batch_size; i++) {
6201                 hash = batch[i]->hash_lock_index;
6202                 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
6203         }
6204         return batch_size;
6205 }
6206
6207 static void raid5_do_work(struct work_struct *work)
6208 {
6209         struct r5worker *worker = container_of(work, struct r5worker, work);
6210         struct r5worker_group *group = worker->group;
6211         struct r5conf *conf = group->conf;
6212         struct mddev *mddev = conf->mddev;
6213         int group_id = group - conf->worker_groups;
6214         int handled;
6215         struct blk_plug plug;
6216
6217         pr_debug("+++ raid5worker active\n");
6218
6219         blk_start_plug(&plug);
6220         handled = 0;
6221         spin_lock_irq(&conf->device_lock);
6222         while (1) {
6223                 int batch_size, released;
6224
6225                 released = release_stripe_list(conf, worker->temp_inactive_list);
6226
6227                 batch_size = handle_active_stripes(conf, group_id, worker,
6228                                                    worker->temp_inactive_list);
6229                 worker->working = false;
6230                 if (!batch_size && !released)
6231                         break;
6232                 handled += batch_size;
6233                 wait_event_lock_irq(mddev->sb_wait,
6234                         !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags),
6235                         conf->device_lock);
6236         }
6237         pr_debug("%d stripes handled\n", handled);
6238
6239         spin_unlock_irq(&conf->device_lock);
6240
6241         flush_deferred_bios(conf);
6242
6243         r5l_flush_stripe_to_raid(conf->log);
6244
6245         async_tx_issue_pending_all();
6246         blk_finish_plug(&plug);
6247
6248         pr_debug("--- raid5worker inactive\n");
6249 }
6250
6251 /*
6252  * This is our raid5 kernel thread.
6253  *
6254  * We scan the hash table for stripes which can be handled now.
6255  * During the scan, completed stripes are saved for us by the interrupt
6256  * handler, so that they will not have to wait for our next wakeup.
6257  */
6258 static void raid5d(struct md_thread *thread)
6259 {
6260         struct mddev *mddev = thread->mddev;
6261         struct r5conf *conf = mddev->private;
6262         int handled;
6263         struct blk_plug plug;
6264
6265         pr_debug("+++ raid5d active\n");
6266
6267         md_check_recovery(mddev);
6268
6269         blk_start_plug(&plug);
6270         handled = 0;
6271         spin_lock_irq(&conf->device_lock);
6272         while (1) {
6273                 struct bio *bio;
6274                 int batch_size, released;
6275                 unsigned int offset;
6276
6277                 released = release_stripe_list(conf, conf->temp_inactive_list);
6278                 if (released)
6279                         clear_bit(R5_DID_ALLOC, &conf->cache_state);
6280
6281                 if (
6282                     !list_empty(&conf->bitmap_list)) {
6283                         /* Now is a good time to flush some bitmap updates */
6284                         conf->seq_flush++;
6285                         spin_unlock_irq(&conf->device_lock);
6286                         bitmap_unplug(mddev->bitmap);
6287                         spin_lock_irq(&conf->device_lock);
6288                         conf->seq_write = conf->seq_flush;
6289                         activate_bit_delay(conf, conf->temp_inactive_list);
6290                 }
6291                 raid5_activate_delayed(conf);
6292
6293                 while ((bio = remove_bio_from_retry(conf, &offset))) {
6294                         int ok;
6295                         spin_unlock_irq(&conf->device_lock);
6296                         ok = retry_aligned_read(conf, bio, offset);
6297                         spin_lock_irq(&conf->device_lock);
6298                         if (!ok)
6299                                 break;
6300                         handled++;
6301                 }
6302
6303                 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
6304                                                    conf->temp_inactive_list);
6305                 if (!batch_size && !released)
6306                         break;
6307                 handled += batch_size;
6308
6309                 if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
6310                         spin_unlock_irq(&conf->device_lock);
6311                         md_check_recovery(mddev);
6312                         spin_lock_irq(&conf->device_lock);
6313                 }
6314         }
6315         pr_debug("%d stripes handled\n", handled);
6316
6317         spin_unlock_irq(&conf->device_lock);
6318         if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
6319             mutex_trylock(&conf->cache_size_mutex)) {
6320                 grow_one_stripe(conf, __GFP_NOWARN);
6321                 /* Set flag even if allocation failed.  This helps
6322                  * slow down allocation requests when mem is short
6323                  */
6324                 set_bit(R5_DID_ALLOC, &conf->cache_state);
6325                 mutex_unlock(&conf->cache_size_mutex);
6326         }
6327
6328         flush_deferred_bios(conf);
6329
6330         r5l_flush_stripe_to_raid(conf->log);
6331
6332         async_tx_issue_pending_all();
6333         blk_finish_plug(&plug);
6334
6335         pr_debug("--- raid5d inactive\n");
6336 }
6337
6338 static ssize_t
6339 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
6340 {
6341         struct r5conf *conf;
6342         int ret = 0;
6343         spin_lock(&mddev->lock);
6344         conf = mddev->private;
6345         if (conf)
6346                 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
6347         spin_unlock(&mddev->lock);
6348         return ret;
6349 }
6350
6351 int
6352 raid5_set_cache_size(struct mddev *mddev, int size)
6353 {
6354         struct r5conf *conf = mddev->private;
6355
6356         if (size <= 16 || size > 32768)
6357                 return -EINVAL;
6358
6359         conf->min_nr_stripes = size;
6360         mutex_lock(&conf->cache_size_mutex);
6361         while (size < conf->max_nr_stripes &&
6362                drop_one_stripe(conf))
6363                 ;
6364         mutex_unlock(&conf->cache_size_mutex);
6365
6366         md_allow_write(mddev);
6367
6368         mutex_lock(&conf->cache_size_mutex);
6369         while (size > conf->max_nr_stripes)
6370                 if (!grow_one_stripe(conf, GFP_KERNEL))
6371                         break;
6372         mutex_unlock(&conf->cache_size_mutex);
6373
6374         return 0;
6375 }
6376 EXPORT_SYMBOL(raid5_set_cache_size);
6377
6378 static ssize_t
6379 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6380 {
6381         struct r5conf *conf;
6382         unsigned long new;
6383         int err;
6384
6385         if (len >= PAGE_SIZE)
6386                 return -EINVAL;
6387         if (kstrtoul(page, 10, &new))
6388                 return -EINVAL;
6389         err = mddev_lock(mddev);
6390         if (err)
6391                 return err;
6392         conf = mddev->private;
6393         if (!conf)
6394                 err = -ENODEV;
6395         else
6396                 err = raid5_set_cache_size(mddev, new);
6397         mddev_unlock(mddev);
6398
6399         return err ?: len;
6400 }
6401
6402 static struct md_sysfs_entry
6403 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6404                                 raid5_show_stripe_cache_size,
6405                                 raid5_store_stripe_cache_size);
6406
6407 static ssize_t
6408 raid5_show_rmw_level(struct mddev  *mddev, char *page)
6409 {
6410         struct r5conf *conf = mddev->private;
6411         if (conf)
6412                 return sprintf(page, "%d\n", conf->rmw_level);
6413         else
6414                 return 0;
6415 }
6416
6417 static ssize_t
6418 raid5_store_rmw_level(struct mddev  *mddev, const char *page, size_t len)
6419 {
6420         struct r5conf *conf = mddev->private;
6421         unsigned long new;
6422
6423         if (!conf)
6424                 return -ENODEV;
6425
6426         if (len >= PAGE_SIZE)
6427                 return -EINVAL;
6428
6429         if (kstrtoul(page, 10, &new))
6430                 return -EINVAL;
6431
6432         if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6433                 return -EINVAL;
6434
6435         if (new != PARITY_DISABLE_RMW &&
6436             new != PARITY_ENABLE_RMW &&
6437             new != PARITY_PREFER_RMW)
6438                 return -EINVAL;
6439
6440         conf->rmw_level = new;
6441         return len;
6442 }
6443
6444 static struct md_sysfs_entry
6445 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6446                          raid5_show_rmw_level,
6447                          raid5_store_rmw_level);
6448
6449
6450 static ssize_t
6451 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6452 {
6453         struct r5conf *conf;
6454         int ret = 0;
6455         spin_lock(&mddev->lock);
6456         conf = mddev->private;
6457         if (conf)
6458                 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6459         spin_unlock(&mddev->lock);
6460         return ret;
6461 }
6462
6463 static ssize_t
6464 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6465 {
6466         struct r5conf *conf;
6467         unsigned long new;
6468         int err;
6469
6470         if (len >= PAGE_SIZE)
6471                 return -EINVAL;
6472         if (kstrtoul(page, 10, &new))
6473                 return -EINVAL;
6474
6475         err = mddev_lock(mddev);
6476         if (err)
6477                 return err;
6478         conf = mddev->private;
6479         if (!conf)
6480                 err = -ENODEV;
6481         else if (new > conf->min_nr_stripes)
6482                 err = -EINVAL;
6483         else
6484                 conf->bypass_threshold = new;
6485         mddev_unlock(mddev);
6486         return err ?: len;
6487 }
6488
6489 static struct md_sysfs_entry
6490 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6491                                         S_IRUGO | S_IWUSR,
6492                                         raid5_show_preread_threshold,
6493                                         raid5_store_preread_threshold);
6494
6495 static ssize_t
6496 raid5_show_skip_copy(struct mddev *mddev, char *page)
6497 {
6498         struct r5conf *conf;
6499         int ret = 0;
6500         spin_lock(&mddev->lock);
6501         conf = mddev->private;
6502         if (conf)
6503                 ret = sprintf(page, "%d\n", conf->skip_copy);
6504         spin_unlock(&mddev->lock);
6505         return ret;
6506 }
6507
6508 static ssize_t
6509 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6510 {
6511         struct r5conf *conf;
6512         unsigned long new;
6513         int err;
6514
6515         if (len >= PAGE_SIZE)
6516                 return -EINVAL;
6517         if (kstrtoul(page, 10, &new))
6518                 return -EINVAL;
6519         new = !!new;
6520
6521         err = mddev_lock(mddev);
6522         if (err)
6523                 return err;
6524         conf = mddev->private;
6525         if (!conf)
6526                 err = -ENODEV;
6527         else if (new != conf->skip_copy) {
6528                 mddev_suspend(mddev);
6529                 conf->skip_copy = new;
6530                 if (new)
6531                         mddev->queue->backing_dev_info->capabilities |=
6532                                 BDI_CAP_STABLE_WRITES;
6533                 else
6534                         mddev->queue->backing_dev_info->capabilities &=
6535                                 ~BDI_CAP_STABLE_WRITES;
6536                 mddev_resume(mddev);
6537         }
6538         mddev_unlock(mddev);
6539         return err ?: len;
6540 }
6541
6542 static struct md_sysfs_entry
6543 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6544                                         raid5_show_skip_copy,
6545                                         raid5_store_skip_copy);
6546
6547 static ssize_t
6548 stripe_cache_active_show(struct mddev *mddev, char *page)
6549 {
6550         struct r5conf *conf = mddev->private;
6551         if (conf)
6552                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6553         else
6554                 return 0;
6555 }
6556
6557 static struct md_sysfs_entry
6558 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6559
6560 static ssize_t
6561 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6562 {
6563         struct r5conf *conf;
6564         int ret = 0;
6565         spin_lock(&mddev->lock);
6566         conf = mddev->private;
6567         if (conf)
6568                 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6569         spin_unlock(&mddev->lock);
6570         return ret;
6571 }
6572
6573 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6574                                int *group_cnt,
6575                                int *worker_cnt_per_group,
6576                                struct r5worker_group **worker_groups);
6577 static ssize_t
6578 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6579 {
6580         struct r5conf *conf;
6581         unsigned int new;
6582         int err;
6583         struct r5worker_group *new_groups, *old_groups;
6584         int group_cnt, worker_cnt_per_group;
6585
6586         if (len >= PAGE_SIZE)
6587                 return -EINVAL;
6588         if (kstrtouint(page, 10, &new))
6589                 return -EINVAL;
6590         /* 8192 should be big enough */
6591         if (new > 8192)
6592                 return -EINVAL;
6593
6594         err = mddev_lock(mddev);
6595         if (err)
6596                 return err;
6597         conf = mddev->private;
6598         if (!conf)
6599                 err = -ENODEV;
6600         else if (new != conf->worker_cnt_per_group) {
6601                 mddev_suspend(mddev);
6602
6603                 old_groups = conf->worker_groups;
6604                 if (old_groups)
6605                         flush_workqueue(raid5_wq);
6606
6607                 err = alloc_thread_groups(conf, new,
6608                                           &group_cnt, &worker_cnt_per_group,
6609                                           &new_groups);
6610                 if (!err) {
6611                         spin_lock_irq(&conf->device_lock);
6612                         conf->group_cnt = group_cnt;
6613                         conf->worker_cnt_per_group = worker_cnt_per_group;
6614                         conf->worker_groups = new_groups;
6615                         spin_unlock_irq(&conf->device_lock);
6616
6617                         if (old_groups)
6618                                 kfree(old_groups[0].workers);
6619                         kfree(old_groups);
6620                 }
6621                 mddev_resume(mddev);
6622         }
6623         mddev_unlock(mddev);
6624
6625         return err ?: len;
6626 }
6627
6628 static struct md_sysfs_entry
6629 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6630                                 raid5_show_group_thread_cnt,
6631                                 raid5_store_group_thread_cnt);
6632
6633 static struct attribute *raid5_attrs[] =  {
6634         &raid5_stripecache_size.attr,
6635         &raid5_stripecache_active.attr,
6636         &raid5_preread_bypass_threshold.attr,
6637         &raid5_group_thread_cnt.attr,
6638         &raid5_skip_copy.attr,
6639         &raid5_rmw_level.attr,
6640         &r5c_journal_mode.attr,
6641         NULL,
6642 };
6643 static struct attribute_group raid5_attrs_group = {
6644         .name = NULL,
6645         .attrs = raid5_attrs,
6646 };
6647
6648 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6649                                int *group_cnt,
6650                                int *worker_cnt_per_group,
6651                                struct r5worker_group **worker_groups)
6652 {
6653         int i, j, k;
6654         ssize_t size;
6655         struct r5worker *workers;
6656
6657         *worker_cnt_per_group = cnt;
6658         if (cnt == 0) {
6659                 *group_cnt = 0;
6660                 *worker_groups = NULL;
6661                 return 0;
6662         }
6663         *group_cnt = num_possible_nodes();
6664         size = sizeof(struct r5worker) * cnt;
6665         workers = kzalloc(size * *group_cnt, GFP_NOIO);
6666         *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6667                                 *group_cnt, GFP_NOIO);
6668         if (!*worker_groups || !workers) {
6669                 kfree(workers);
6670                 kfree(*worker_groups);
6671                 return -ENOMEM;
6672         }
6673
6674         for (i = 0; i < *group_cnt; i++) {
6675                 struct r5worker_group *group;
6676
6677                 group = &(*worker_groups)[i];
6678                 INIT_LIST_HEAD(&group->handle_list);
6679                 INIT_LIST_HEAD(&group->loprio_list);
6680                 group->conf = conf;
6681                 group->workers = workers + i * cnt;
6682
6683                 for (j = 0; j < cnt; j++) {
6684                         struct r5worker *worker = group->workers + j;
6685                         worker->group = group;
6686                         INIT_WORK(&worker->work, raid5_do_work);
6687
6688                         for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6689                                 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6690                 }
6691         }
6692
6693         return 0;
6694 }
6695
6696 static void free_thread_groups(struct r5conf *conf)
6697 {
6698         if (conf->worker_groups)
6699                 kfree(conf->worker_groups[0].workers);
6700         kfree(conf->worker_groups);
6701         conf->worker_groups = NULL;
6702 }
6703
6704 static sector_t
6705 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6706 {
6707         struct r5conf *conf = mddev->private;
6708
6709         if (!sectors)
6710                 sectors = mddev->dev_sectors;
6711         if (!raid_disks)
6712                 /* size is defined by the smallest of previous and new size */
6713                 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6714
6715         sectors &= ~((sector_t)conf->chunk_sectors - 1);
6716         sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6717         return sectors * (raid_disks - conf->max_degraded);
6718 }
6719
6720 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6721 {
6722         safe_put_page(percpu->spare_page);
6723         if (percpu->scribble)
6724                 flex_array_free(percpu->scribble);
6725         percpu->spare_page = NULL;
6726         percpu->scribble = NULL;
6727 }
6728
6729 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6730 {
6731         if (conf->level == 6 && !percpu->spare_page)
6732                 percpu->spare_page = alloc_page(GFP_KERNEL);
6733         if (!percpu->scribble)
6734                 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6735                                                       conf->previous_raid_disks),
6736                                                   max(conf->chunk_sectors,
6737                                                       conf->prev_chunk_sectors)
6738                                                    / STRIPE_SECTORS,
6739                                                   GFP_KERNEL);
6740
6741         if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6742                 free_scratch_buffer(conf, percpu);
6743                 return -ENOMEM;
6744         }
6745
6746         return 0;
6747 }
6748
6749 static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
6750 {
6751         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6752
6753         free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6754         return 0;
6755 }
6756
6757 static void raid5_free_percpu(struct r5conf *conf)
6758 {
6759         if (!conf->percpu)
6760                 return;
6761
6762         cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6763         free_percpu(conf->percpu);
6764 }
6765
6766 static void free_conf(struct r5conf *conf)
6767 {
6768         int i;
6769
6770         log_exit(conf);
6771
6772         if (conf->shrinker.nr_deferred)
6773                 unregister_shrinker(&conf->shrinker);
6774
6775         free_thread_groups(conf);
6776         shrink_stripes(conf);
6777         raid5_free_percpu(conf);
6778         for (i = 0; i < conf->pool_size; i++)
6779                 if (conf->disks[i].extra_page)
6780                         put_page(conf->disks[i].extra_page);
6781         kfree(conf->disks);
6782         if (conf->bio_split)
6783                 bioset_free(conf->bio_split);
6784         kfree(conf->stripe_hashtbl);
6785         kfree(conf->pending_data);
6786         kfree(conf);
6787 }
6788
6789 static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
6790 {
6791         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6792         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6793
6794         if (alloc_scratch_buffer(conf, percpu)) {
6795                 pr_warn("%s: failed memory allocation for cpu%u\n",
6796                         __func__, cpu);
6797                 return -ENOMEM;
6798         }
6799         return 0;
6800 }
6801
6802 static int raid5_alloc_percpu(struct r5conf *conf)
6803 {
6804         int err = 0;
6805
6806         conf->percpu = alloc_percpu(struct raid5_percpu);
6807         if (!conf->percpu)
6808                 return -ENOMEM;
6809
6810         err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6811         if (!err) {
6812                 conf->scribble_disks = max(conf->raid_disks,
6813                         conf->previous_raid_disks);
6814                 conf->scribble_sectors = max(conf->chunk_sectors,
6815                         conf->prev_chunk_sectors);
6816         }
6817         return err;
6818 }
6819
6820 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6821                                       struct shrink_control *sc)
6822 {
6823         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6824         unsigned long ret = SHRINK_STOP;
6825
6826         if (mutex_trylock(&conf->cache_size_mutex)) {
6827                 ret= 0;
6828                 while (ret < sc->nr_to_scan &&
6829                        conf->max_nr_stripes > conf->min_nr_stripes) {
6830                         if (drop_one_stripe(conf) == 0) {
6831                                 ret = SHRINK_STOP;
6832                                 break;
6833                         }
6834                         ret++;
6835                 }
6836                 mutex_unlock(&conf->cache_size_mutex);
6837         }
6838         return ret;
6839 }
6840
6841 static unsigned long raid5_cache_count(struct shrinker *shrink,
6842                                        struct shrink_control *sc)
6843 {
6844         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6845
6846         if (conf->max_nr_stripes < conf->min_nr_stripes)
6847                 /* unlikely, but not impossible */
6848                 return 0;
6849         return conf->max_nr_stripes - conf->min_nr_stripes;
6850 }
6851
6852 static struct r5conf *setup_conf(struct mddev *mddev)
6853 {
6854         struct r5conf *conf;
6855         int raid_disk, memory, max_disks;
6856         struct md_rdev *rdev;
6857         struct disk_info *disk;
6858         char pers_name[6];
6859         int i;
6860         int group_cnt, worker_cnt_per_group;
6861         struct r5worker_group *new_group;
6862
6863         if (mddev->new_level != 5
6864             && mddev->new_level != 4
6865             && mddev->new_level != 6) {
6866                 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6867                         mdname(mddev), mddev->new_level);
6868                 return ERR_PTR(-EIO);
6869         }
6870         if ((mddev->new_level == 5
6871              && !algorithm_valid_raid5(mddev->new_layout)) ||
6872             (mddev->new_level == 6
6873              && !algorithm_valid_raid6(mddev->new_layout))) {
6874                 pr_warn("md/raid:%s: layout %d not supported\n",
6875                         mdname(mddev), mddev->new_layout);
6876                 return ERR_PTR(-EIO);
6877         }
6878         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6879                 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6880                         mdname(mddev), mddev->raid_disks);
6881                 return ERR_PTR(-EINVAL);
6882         }
6883
6884         if (!mddev->new_chunk_sectors ||
6885             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6886             !is_power_of_2(mddev->new_chunk_sectors)) {
6887                 pr_warn("md/raid:%s: invalid chunk size %d\n",
6888                         mdname(mddev), mddev->new_chunk_sectors << 9);
6889                 return ERR_PTR(-EINVAL);
6890         }
6891
6892         conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6893         if (conf == NULL)
6894                 goto abort;
6895         INIT_LIST_HEAD(&conf->free_list);
6896         INIT_LIST_HEAD(&conf->pending_list);
6897         conf->pending_data = kzalloc(sizeof(struct r5pending_data) *
6898                 PENDING_IO_MAX, GFP_KERNEL);
6899         if (!conf->pending_data)
6900                 goto abort;
6901         for (i = 0; i < PENDING_IO_MAX; i++)
6902                 list_add(&conf->pending_data[i].sibling, &conf->free_list);
6903         /* Don't enable multi-threading by default*/
6904         if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6905                                  &new_group)) {
6906                 conf->group_cnt = group_cnt;
6907                 conf->worker_cnt_per_group = worker_cnt_per_group;
6908                 conf->worker_groups = new_group;
6909         } else
6910                 goto abort;
6911         spin_lock_init(&conf->device_lock);
6912         seqcount_init(&conf->gen_lock);
6913         mutex_init(&conf->cache_size_mutex);
6914         init_waitqueue_head(&conf->wait_for_quiescent);
6915         init_waitqueue_head(&conf->wait_for_stripe);
6916         init_waitqueue_head(&conf->wait_for_overlap);
6917         INIT_LIST_HEAD(&conf->handle_list);
6918         INIT_LIST_HEAD(&conf->loprio_list);
6919         INIT_LIST_HEAD(&conf->hold_list);
6920         INIT_LIST_HEAD(&conf->delayed_list);
6921         INIT_LIST_HEAD(&conf->bitmap_list);
6922         init_llist_head(&conf->released_stripes);
6923         atomic_set(&conf->active_stripes, 0);
6924         atomic_set(&conf->preread_active_stripes, 0);
6925         atomic_set(&conf->active_aligned_reads, 0);
6926         spin_lock_init(&conf->pending_bios_lock);
6927         conf->batch_bio_dispatch = true;
6928         rdev_for_each(rdev, mddev) {
6929                 if (test_bit(Journal, &rdev->flags))
6930                         continue;
6931                 if (blk_queue_nonrot(bdev_get_queue(rdev->bdev))) {
6932                         conf->batch_bio_dispatch = false;
6933                         break;
6934                 }
6935         }
6936
6937         conf->bypass_threshold = BYPASS_THRESHOLD;
6938         conf->recovery_disabled = mddev->recovery_disabled - 1;
6939
6940         conf->raid_disks = mddev->raid_disks;
6941         if (mddev->reshape_position == MaxSector)
6942                 conf->previous_raid_disks = mddev->raid_disks;
6943         else
6944                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6945         max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6946
6947         conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6948                               GFP_KERNEL);
6949
6950         if (!conf->disks)
6951                 goto abort;
6952
6953         for (i = 0; i < max_disks; i++) {
6954                 conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
6955                 if (!conf->disks[i].extra_page)
6956                         goto abort;
6957         }
6958
6959         conf->bio_split = bioset_create(BIO_POOL_SIZE, 0, 0);
6960         if (!conf->bio_split)
6961                 goto abort;
6962         conf->mddev = mddev;
6963
6964         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6965                 goto abort;
6966
6967         /* We init hash_locks[0] separately to that it can be used
6968          * as the reference lock in the spin_lock_nest_lock() call
6969          * in lock_all_device_hash_locks_irq in order to convince
6970          * lockdep that we know what we are doing.
6971          */
6972         spin_lock_init(conf->hash_locks);
6973         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6974                 spin_lock_init(conf->hash_locks + i);
6975
6976         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6977                 INIT_LIST_HEAD(conf->inactive_list + i);
6978
6979         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6980                 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6981
6982         atomic_set(&conf->r5c_cached_full_stripes, 0);
6983         INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
6984         atomic_set(&conf->r5c_cached_partial_stripes, 0);
6985         INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
6986         atomic_set(&conf->r5c_flushing_full_stripes, 0);
6987         atomic_set(&conf->r5c_flushing_partial_stripes, 0);
6988
6989         conf->level = mddev->new_level;
6990         conf->chunk_sectors = mddev->new_chunk_sectors;
6991         if (raid5_alloc_percpu(conf) != 0)
6992                 goto abort;
6993
6994         pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6995
6996         rdev_for_each(rdev, mddev) {
6997                 raid_disk = rdev->raid_disk;
6998                 if (raid_disk >= max_disks
6999                     || raid_disk < 0 || test_bit(Journal, &rdev->flags))
7000                         continue;
7001                 disk = conf->disks + raid_disk;
7002
7003                 if (test_bit(Replacement, &rdev->flags)) {
7004                         if (disk->replacement)
7005                                 goto abort;
7006                         disk->replacement = rdev;
7007                 } else {
7008                         if (disk->rdev)
7009                                 goto abort;
7010                         disk->rdev = rdev;
7011                 }
7012
7013                 if (test_bit(In_sync, &rdev->flags)) {
7014                         char b[BDEVNAME_SIZE];
7015                         pr_info("md/raid:%s: device %s operational as raid disk %d\n",
7016                                 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
7017                 } else if (rdev->saved_raid_disk != raid_disk)
7018                         /* Cannot rely on bitmap to complete recovery */
7019                         conf->fullsync = 1;
7020         }
7021
7022         conf->level = mddev->new_level;
7023         if (conf->level == 6) {
7024                 conf->max_degraded = 2;
7025                 if (raid6_call.xor_syndrome)
7026                         conf->rmw_level = PARITY_ENABLE_RMW;
7027                 else
7028                         conf->rmw_level = PARITY_DISABLE_RMW;
7029         } else {
7030                 conf->max_degraded = 1;
7031                 conf->rmw_level = PARITY_ENABLE_RMW;
7032         }
7033         conf->algorithm = mddev->new_layout;
7034         conf->reshape_progress = mddev->reshape_position;
7035         if (conf->reshape_progress != MaxSector) {
7036                 conf->prev_chunk_sectors = mddev->chunk_sectors;
7037                 conf->prev_algo = mddev->layout;
7038         } else {
7039                 conf->prev_chunk_sectors = conf->chunk_sectors;
7040                 conf->prev_algo = conf->algorithm;
7041         }
7042
7043         conf->min_nr_stripes = NR_STRIPES;
7044         if (mddev->reshape_position != MaxSector) {
7045                 int stripes = max_t(int,
7046                         ((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4,
7047                         ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4);
7048                 conf->min_nr_stripes = max(NR_STRIPES, stripes);
7049                 if (conf->min_nr_stripes != NR_STRIPES)
7050                         pr_info("md/raid:%s: force stripe size %d for reshape\n",
7051                                 mdname(mddev), conf->min_nr_stripes);
7052         }
7053         memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
7054                  max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
7055         atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
7056         if (grow_stripes(conf, conf->min_nr_stripes)) {
7057                 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7058                         mdname(mddev), memory);
7059                 goto abort;
7060         } else
7061                 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
7062         /*
7063          * Losing a stripe head costs more than the time to refill it,
7064          * it reduces the queue depth and so can hurt throughput.
7065          * So set it rather large, scaled by number of devices.
7066          */
7067         conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
7068         conf->shrinker.scan_objects = raid5_cache_scan;
7069         conf->shrinker.count_objects = raid5_cache_count;
7070         conf->shrinker.batch = 128;
7071         conf->shrinker.flags = 0;
7072         if (register_shrinker(&conf->shrinker)) {
7073                 pr_warn("md/raid:%s: couldn't register shrinker.\n",
7074                         mdname(mddev));
7075                 goto abort;
7076         }
7077
7078         sprintf(pers_name, "raid%d", mddev->new_level);
7079         conf->thread = md_register_thread(raid5d, mddev, pers_name);
7080         if (!conf->thread) {
7081                 pr_warn("md/raid:%s: couldn't allocate thread.\n",
7082                         mdname(mddev));
7083                 goto abort;
7084         }
7085
7086         return conf;
7087
7088  abort:
7089         if (conf) {
7090                 free_conf(conf);
7091                 return ERR_PTR(-EIO);
7092         } else
7093                 return ERR_PTR(-ENOMEM);
7094 }
7095
7096 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
7097 {
7098         switch (algo) {
7099         case ALGORITHM_PARITY_0:
7100                 if (raid_disk < max_degraded)
7101                         return 1;
7102                 break;
7103         case ALGORITHM_PARITY_N:
7104                 if (raid_disk >= raid_disks - max_degraded)
7105                         return 1;
7106                 break;
7107         case ALGORITHM_PARITY_0_6:
7108                 if (raid_disk == 0 ||
7109                     raid_disk == raid_disks - 1)
7110                         return 1;
7111                 break;
7112         case ALGORITHM_LEFT_ASYMMETRIC_6:
7113         case ALGORITHM_RIGHT_ASYMMETRIC_6:
7114         case ALGORITHM_LEFT_SYMMETRIC_6:
7115         case ALGORITHM_RIGHT_SYMMETRIC_6:
7116                 if (raid_disk == raid_disks - 1)
7117                         return 1;
7118         }
7119         return 0;
7120 }
7121
7122 static int raid5_run(struct mddev *mddev)
7123 {
7124         struct r5conf *conf;
7125         int working_disks = 0;
7126         int dirty_parity_disks = 0;
7127         struct md_rdev *rdev;
7128         struct md_rdev *journal_dev = NULL;
7129         sector_t reshape_offset = 0;
7130         int i;
7131         long long min_offset_diff = 0;
7132         int first = 1;
7133
7134         if (mddev_init_writes_pending(mddev) < 0)
7135                 return -ENOMEM;
7136
7137         if (mddev->recovery_cp != MaxSector)
7138                 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7139                           mdname(mddev));
7140
7141         rdev_for_each(rdev, mddev) {
7142                 long long diff;
7143
7144                 if (test_bit(Journal, &rdev->flags)) {
7145                         journal_dev = rdev;
7146                         continue;
7147                 }
7148                 if (rdev->raid_disk < 0)
7149                         continue;
7150                 diff = (rdev->new_data_offset - rdev->data_offset);
7151                 if (first) {
7152                         min_offset_diff = diff;
7153                         first = 0;
7154                 } else if (mddev->reshape_backwards &&
7155                          diff < min_offset_diff)
7156                         min_offset_diff = diff;
7157                 else if (!mddev->reshape_backwards &&
7158                          diff > min_offset_diff)
7159                         min_offset_diff = diff;
7160         }
7161
7162         if ((test_bit(MD_HAS_JOURNAL, &mddev->flags) || journal_dev) &&
7163             (mddev->bitmap_info.offset || mddev->bitmap_info.file)) {
7164                 pr_notice("md/raid:%s: array cannot have both journal and bitmap\n",
7165                           mdname(mddev));
7166                 return -EINVAL;
7167         }
7168
7169         if (mddev->reshape_position != MaxSector) {
7170                 /* Check that we can continue the reshape.
7171                  * Difficulties arise if the stripe we would write to
7172                  * next is at or after the stripe we would read from next.
7173                  * For a reshape that changes the number of devices, this
7174                  * is only possible for a very short time, and mdadm makes
7175                  * sure that time appears to have past before assembling
7176                  * the array.  So we fail if that time hasn't passed.
7177                  * For a reshape that keeps the number of devices the same
7178                  * mdadm must be monitoring the reshape can keeping the
7179                  * critical areas read-only and backed up.  It will start
7180                  * the array in read-only mode, so we check for that.
7181                  */
7182                 sector_t here_new, here_old;
7183                 int old_disks;
7184                 int max_degraded = (mddev->level == 6 ? 2 : 1);
7185                 int chunk_sectors;
7186                 int new_data_disks;
7187
7188                 if (journal_dev) {
7189                         pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7190                                 mdname(mddev));
7191                         return -EINVAL;
7192                 }
7193
7194                 if (mddev->new_level != mddev->level) {
7195                         pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7196                                 mdname(mddev));
7197                         return -EINVAL;
7198                 }
7199                 old_disks = mddev->raid_disks - mddev->delta_disks;
7200                 /* reshape_position must be on a new-stripe boundary, and one
7201                  * further up in new geometry must map after here in old
7202                  * geometry.
7203                  * If the chunk sizes are different, then as we perform reshape
7204                  * in units of the largest of the two, reshape_position needs
7205                  * be a multiple of the largest chunk size times new data disks.
7206                  */
7207                 here_new = mddev->reshape_position;
7208                 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
7209                 new_data_disks = mddev->raid_disks - max_degraded;
7210                 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
7211                         pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7212                                 mdname(mddev));
7213                         return -EINVAL;
7214                 }
7215                 reshape_offset = here_new * chunk_sectors;
7216                 /* here_new is the stripe we will write to */
7217                 here_old = mddev->reshape_position;
7218                 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
7219                 /* here_old is the first stripe that we might need to read
7220                  * from */
7221                 if (mddev->delta_disks == 0) {
7222                         /* We cannot be sure it is safe to start an in-place
7223                          * reshape.  It is only safe if user-space is monitoring
7224                          * and taking constant backups.
7225                          * mdadm always starts a situation like this in
7226                          * readonly mode so it can take control before
7227                          * allowing any writes.  So just check for that.
7228                          */
7229                         if (abs(min_offset_diff) >= mddev->chunk_sectors &&
7230                             abs(min_offset_diff) >= mddev->new_chunk_sectors)
7231                                 /* not really in-place - so OK */;
7232                         else if (mddev->ro == 0) {
7233                                 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7234                                         mdname(mddev));
7235                                 return -EINVAL;
7236                         }
7237                 } else if (mddev->reshape_backwards
7238                     ? (here_new * chunk_sectors + min_offset_diff <=
7239                        here_old * chunk_sectors)
7240                     : (here_new * chunk_sectors >=
7241                        here_old * chunk_sectors + (-min_offset_diff))) {
7242                         /* Reading from the same stripe as writing to - bad */
7243                         pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7244                                 mdname(mddev));
7245                         return -EINVAL;
7246                 }
7247                 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
7248                 /* OK, we should be able to continue; */
7249         } else {
7250                 BUG_ON(mddev->level != mddev->new_level);
7251                 BUG_ON(mddev->layout != mddev->new_layout);
7252                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
7253                 BUG_ON(mddev->delta_disks != 0);
7254         }
7255
7256         if (test_bit(MD_HAS_JOURNAL, &mddev->flags) &&
7257             test_bit(MD_HAS_PPL, &mddev->flags)) {
7258                 pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7259                         mdname(mddev));
7260                 clear_bit(MD_HAS_PPL, &mddev->flags);
7261                 clear_bit(MD_HAS_MULTIPLE_PPLS, &mddev->flags);
7262         }
7263
7264         if (mddev->private == NULL)
7265                 conf = setup_conf(mddev);
7266         else
7267                 conf = mddev->private;
7268
7269         if (IS_ERR(conf))
7270                 return PTR_ERR(conf);
7271
7272         if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
7273                 if (!journal_dev) {
7274                         pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7275                                 mdname(mddev));
7276                         mddev->ro = 1;
7277                         set_disk_ro(mddev->gendisk, 1);
7278                 } else if (mddev->recovery_cp == MaxSector)
7279                         set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
7280         }
7281
7282         conf->min_offset_diff = min_offset_diff;
7283         mddev->thread = conf->thread;
7284         conf->thread = NULL;
7285         mddev->private = conf;
7286
7287         for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
7288              i++) {
7289                 rdev = conf->disks[i].rdev;
7290                 if (!rdev && conf->disks[i].replacement) {
7291                         /* The replacement is all we have yet */
7292                         rdev = conf->disks[i].replacement;
7293                         conf->disks[i].replacement = NULL;
7294                         clear_bit(Replacement, &rdev->flags);
7295                         conf->disks[i].rdev = rdev;
7296                 }
7297                 if (!rdev)
7298                         continue;
7299                 if (conf->disks[i].replacement &&
7300                     conf->reshape_progress != MaxSector) {
7301                         /* replacements and reshape simply do not mix. */
7302                         pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7303                         goto abort;
7304                 }
7305                 if (test_bit(In_sync, &rdev->flags)) {
7306                         working_disks++;
7307                         continue;
7308                 }
7309                 /* This disc is not fully in-sync.  However if it
7310                  * just stored parity (beyond the recovery_offset),
7311                  * when we don't need to be concerned about the
7312                  * array being dirty.
7313                  * When reshape goes 'backwards', we never have
7314                  * partially completed devices, so we only need
7315                  * to worry about reshape going forwards.
7316                  */
7317                 /* Hack because v0.91 doesn't store recovery_offset properly. */
7318                 if (mddev->major_version == 0 &&
7319                     mddev->minor_version > 90)
7320                         rdev->recovery_offset = reshape_offset;
7321
7322                 if (rdev->recovery_offset < reshape_offset) {
7323                         /* We need to check old and new layout */
7324                         if (!only_parity(rdev->raid_disk,
7325                                          conf->algorithm,
7326                                          conf->raid_disks,
7327                                          conf->max_degraded))
7328                                 continue;
7329                 }
7330                 if (!only_parity(rdev->raid_disk,
7331                                  conf->prev_algo,
7332                                  conf->previous_raid_disks,
7333                                  conf->max_degraded))
7334                         continue;
7335                 dirty_parity_disks++;
7336         }
7337
7338         /*
7339          * 0 for a fully functional array, 1 or 2 for a degraded array.
7340          */
7341         mddev->degraded = raid5_calc_degraded(conf);
7342
7343         if (has_failed(conf)) {
7344                 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7345                         mdname(mddev), mddev->degraded, conf->raid_disks);
7346                 goto abort;
7347         }
7348
7349         /* device size must be a multiple of chunk size */
7350         mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
7351         mddev->resync_max_sectors = mddev->dev_sectors;
7352
7353         if (mddev->degraded > dirty_parity_disks &&
7354             mddev->recovery_cp != MaxSector) {
7355                 if (test_bit(MD_HAS_PPL, &mddev->flags))
7356                         pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7357                                 mdname(mddev));
7358                 else if (mddev->ok_start_degraded)
7359                         pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7360                                 mdname(mddev));
7361                 else {
7362                         pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7363                                 mdname(mddev));
7364                         goto abort;
7365                 }
7366         }
7367
7368         pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7369                 mdname(mddev), conf->level,
7370                 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
7371                 mddev->new_layout);
7372
7373         print_raid5_conf(conf);
7374
7375         if (conf->reshape_progress != MaxSector) {
7376                 conf->reshape_safe = conf->reshape_progress;
7377                 atomic_set(&conf->reshape_stripes, 0);
7378                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7379                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7380                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7381                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7382                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7383                                                         "reshape");
7384         }
7385
7386         /* Ok, everything is just fine now */
7387         if (mddev->to_remove == &raid5_attrs_group)
7388                 mddev->to_remove = NULL;
7389         else if (mddev->kobj.sd &&
7390             sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
7391                 pr_warn("raid5: failed to create sysfs attributes for %s\n",
7392                         mdname(mddev));
7393         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7394
7395         if (mddev->queue) {
7396                 int chunk_size;
7397                 /* read-ahead size must cover two whole stripes, which
7398                  * is 2 * (datadisks) * chunksize where 'n' is the
7399                  * number of raid devices
7400                  */
7401                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
7402                 int stripe = data_disks *
7403                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
7404                 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
7405                         mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
7406
7407                 chunk_size = mddev->chunk_sectors << 9;
7408                 blk_queue_io_min(mddev->queue, chunk_size);
7409                 blk_queue_io_opt(mddev->queue, chunk_size *
7410                                  (conf->raid_disks - conf->max_degraded));
7411                 mddev->queue->limits.raid_partial_stripes_expensive = 1;
7412                 /*
7413                  * We can only discard a whole stripe. It doesn't make sense to
7414                  * discard data disk but write parity disk
7415                  */
7416                 stripe = stripe * PAGE_SIZE;
7417                 /* Round up to power of 2, as discard handling
7418                  * currently assumes that */
7419                 while ((stripe-1) & stripe)
7420                         stripe = (stripe | (stripe-1)) + 1;
7421                 mddev->queue->limits.discard_alignment = stripe;
7422                 mddev->queue->limits.discard_granularity = stripe;
7423
7424                 blk_queue_max_write_same_sectors(mddev->queue, 0);
7425                 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
7426
7427                 rdev_for_each(rdev, mddev) {
7428                         disk_stack_limits(mddev->gendisk, rdev->bdev,
7429                                           rdev->data_offset << 9);
7430                         disk_stack_limits(mddev->gendisk, rdev->bdev,
7431                                           rdev->new_data_offset << 9);
7432                 }
7433
7434                 /*
7435                  * zeroing is required, otherwise data
7436                  * could be lost. Consider a scenario: discard a stripe
7437                  * (the stripe could be inconsistent if
7438                  * discard_zeroes_data is 0); write one disk of the
7439                  * stripe (the stripe could be inconsistent again
7440                  * depending on which disks are used to calculate
7441                  * parity); the disk is broken; The stripe data of this
7442                  * disk is lost.
7443                  *
7444                  * We only allow DISCARD if the sysadmin has confirmed that
7445                  * only safe devices are in use by setting a module parameter.
7446                  * A better idea might be to turn DISCARD into WRITE_ZEROES
7447                  * requests, as that is required to be safe.
7448                  */
7449                 if (devices_handle_discard_safely &&
7450                     mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7451                     mddev->queue->limits.discard_granularity >= stripe)
7452                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
7453                                                 mddev->queue);
7454                 else
7455                         queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
7456                                                 mddev->queue);
7457
7458                 blk_queue_max_hw_sectors(mddev->queue, UINT_MAX);
7459         }
7460
7461         if (log_init(conf, journal_dev, raid5_has_ppl(conf)))
7462                 goto abort;
7463
7464         return 0;
7465 abort:
7466         md_unregister_thread(&mddev->thread);
7467         print_raid5_conf(conf);
7468         free_conf(conf);
7469         mddev->private = NULL;
7470         pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
7471         return -EIO;
7472 }
7473
7474 static void raid5_free(struct mddev *mddev, void *priv)
7475 {
7476         struct r5conf *conf = priv;
7477
7478         free_conf(conf);
7479         mddev->to_remove = &raid5_attrs_group;
7480 }
7481
7482 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
7483 {
7484         struct r5conf *conf = mddev->private;
7485         int i;
7486
7487         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7488                 conf->chunk_sectors / 2, mddev->layout);
7489         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7490         rcu_read_lock();
7491         for (i = 0; i < conf->raid_disks; i++) {
7492                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
7493                 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
7494         }
7495         rcu_read_unlock();
7496         seq_printf (seq, "]");
7497 }
7498
7499 static void print_raid5_conf (struct r5conf *conf)
7500 {
7501         int i;
7502         struct disk_info *tmp;
7503
7504         pr_debug("RAID conf printout:\n");
7505         if (!conf) {
7506                 pr_debug("(conf==NULL)\n");
7507                 return;
7508         }
7509         pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
7510                conf->raid_disks,
7511                conf->raid_disks - conf->mddev->degraded);
7512
7513         for (i = 0; i < conf->raid_disks; i++) {
7514                 char b[BDEVNAME_SIZE];
7515                 tmp = conf->disks + i;
7516                 if (tmp->rdev)
7517                         pr_debug(" disk %d, o:%d, dev:%s\n",
7518                                i, !test_bit(Faulty, &tmp->rdev->flags),
7519                                bdevname(tmp->rdev->bdev, b));
7520         }
7521 }
7522
7523 static int raid5_spare_active(struct mddev *mddev)
7524 {
7525         int i;
7526         struct r5conf *conf = mddev->private;
7527         struct disk_info *tmp;
7528         int count = 0;
7529         unsigned long flags;
7530
7531         for (i = 0; i < conf->raid_disks; i++) {
7532                 tmp = conf->disks + i;
7533                 if (tmp->replacement
7534                     && tmp->replacement->recovery_offset == MaxSector
7535                     && !test_bit(Faulty, &tmp->replacement->flags)
7536                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7537                         /* Replacement has just become active. */
7538                         if (!tmp->rdev
7539                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7540                                 count++;
7541                         if (tmp->rdev) {
7542                                 /* Replaced device not technically faulty,
7543                                  * but we need to be sure it gets removed
7544                                  * and never re-added.
7545                                  */
7546                                 set_bit(Faulty, &tmp->rdev->flags);
7547                                 sysfs_notify_dirent_safe(
7548                                         tmp->rdev->sysfs_state);
7549                         }
7550                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7551                 } else if (tmp->rdev
7552                     && tmp->rdev->recovery_offset == MaxSector
7553                     && !test_bit(Faulty, &tmp->rdev->flags)
7554                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7555                         count++;
7556                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7557                 }
7558         }
7559         spin_lock_irqsave(&conf->device_lock, flags);
7560         mddev->degraded = raid5_calc_degraded(conf);
7561         spin_unlock_irqrestore(&conf->device_lock, flags);
7562         print_raid5_conf(conf);
7563         return count;
7564 }
7565
7566 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7567 {
7568         struct r5conf *conf = mddev->private;
7569         int err = 0;
7570         int number = rdev->raid_disk;
7571         struct md_rdev **rdevp;
7572         struct disk_info *p = conf->disks + number;
7573
7574         print_raid5_conf(conf);
7575         if (test_bit(Journal, &rdev->flags) && conf->log) {
7576                 /*
7577                  * we can't wait pending write here, as this is called in
7578                  * raid5d, wait will deadlock.
7579                  * neilb: there is no locking about new writes here,
7580                  * so this cannot be safe.
7581                  */
7582                 if (atomic_read(&conf->active_stripes) ||
7583                     atomic_read(&conf->r5c_cached_full_stripes) ||
7584                     atomic_read(&conf->r5c_cached_partial_stripes)) {
7585                         return -EBUSY;
7586                 }
7587                 log_exit(conf);
7588                 return 0;
7589         }
7590         if (rdev == p->rdev)
7591                 rdevp = &p->rdev;
7592         else if (rdev == p->replacement)
7593                 rdevp = &p->replacement;
7594         else
7595                 return 0;
7596
7597         if (number >= conf->raid_disks &&
7598             conf->reshape_progress == MaxSector)
7599                 clear_bit(In_sync, &rdev->flags);
7600
7601         if (test_bit(In_sync, &rdev->flags) ||
7602             atomic_read(&rdev->nr_pending)) {
7603                 err = -EBUSY;
7604                 goto abort;
7605         }
7606         /* Only remove non-faulty devices if recovery
7607          * isn't possible.
7608          */
7609         if (!test_bit(Faulty, &rdev->flags) &&
7610             mddev->recovery_disabled != conf->recovery_disabled &&
7611             !has_failed(conf) &&
7612             (!p->replacement || p->replacement == rdev) &&
7613             number < conf->raid_disks) {
7614                 err = -EBUSY;
7615                 goto abort;
7616         }
7617         *rdevp = NULL;
7618         if (!test_bit(RemoveSynchronized, &rdev->flags)) {
7619                 synchronize_rcu();
7620                 if (atomic_read(&rdev->nr_pending)) {
7621                         /* lost the race, try later */
7622                         err = -EBUSY;
7623                         *rdevp = rdev;
7624                 }
7625         }
7626         if (!err) {
7627                 err = log_modify(conf, rdev, false);
7628                 if (err)
7629                         goto abort;
7630         }
7631         if (p->replacement) {
7632                 /* We must have just cleared 'rdev' */
7633                 p->rdev = p->replacement;
7634                 clear_bit(Replacement, &p->replacement->flags);
7635                 smp_mb(); /* Make sure other CPUs may see both as identical
7636                            * but will never see neither - if they are careful
7637                            */
7638                 p->replacement = NULL;
7639
7640                 if (!err)
7641                         err = log_modify(conf, p->rdev, true);
7642         }
7643
7644         clear_bit(WantReplacement, &rdev->flags);
7645 abort:
7646
7647         print_raid5_conf(conf);
7648         return err;
7649 }
7650
7651 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7652 {
7653         struct r5conf *conf = mddev->private;
7654         int err = -EEXIST;
7655         int disk;
7656         struct disk_info *p;
7657         int first = 0;
7658         int last = conf->raid_disks - 1;
7659
7660         if (test_bit(Journal, &rdev->flags)) {
7661                 if (conf->log)
7662                         return -EBUSY;
7663
7664                 rdev->raid_disk = 0;
7665                 /*
7666                  * The array is in readonly mode if journal is missing, so no
7667                  * write requests running. We should be safe
7668                  */
7669                 log_init(conf, rdev, false);
7670                 return 0;
7671         }
7672         if (mddev->recovery_disabled == conf->recovery_disabled)
7673                 return -EBUSY;
7674
7675         if (rdev->saved_raid_disk < 0 && has_failed(conf))
7676                 /* no point adding a device */
7677                 return -EINVAL;
7678
7679         if (rdev->raid_disk >= 0)
7680                 first = last = rdev->raid_disk;
7681
7682         /*
7683          * find the disk ... but prefer rdev->saved_raid_disk
7684          * if possible.
7685          */
7686         if (rdev->saved_raid_disk >= 0 &&
7687             rdev->saved_raid_disk >= first &&
7688             conf->disks[rdev->saved_raid_disk].rdev == NULL)
7689                 first = rdev->saved_raid_disk;
7690
7691         for (disk = first; disk <= last; disk++) {
7692                 p = conf->disks + disk;
7693                 if (p->rdev == NULL) {
7694                         clear_bit(In_sync, &rdev->flags);
7695                         rdev->raid_disk = disk;
7696                         if (rdev->saved_raid_disk != disk)
7697                                 conf->fullsync = 1;
7698                         rcu_assign_pointer(p->rdev, rdev);
7699
7700                         err = log_modify(conf, rdev, true);
7701
7702                         goto out;
7703                 }
7704         }
7705         for (disk = first; disk <= last; disk++) {
7706                 p = conf->disks + disk;
7707                 if (test_bit(WantReplacement, &p->rdev->flags) &&
7708                     p->replacement == NULL) {
7709                         clear_bit(In_sync, &rdev->flags);
7710                         set_bit(Replacement, &rdev->flags);
7711                         rdev->raid_disk = disk;
7712                         err = 0;
7713                         conf->fullsync = 1;
7714                         rcu_assign_pointer(p->replacement, rdev);
7715                         break;
7716                 }
7717         }
7718 out:
7719         print_raid5_conf(conf);
7720         return err;
7721 }
7722
7723 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7724 {
7725         /* no resync is happening, and there is enough space
7726          * on all devices, so we can resize.
7727          * We need to make sure resync covers any new space.
7728          * If the array is shrinking we should possibly wait until
7729          * any io in the removed space completes, but it hardly seems
7730          * worth it.
7731          */
7732         sector_t newsize;
7733         struct r5conf *conf = mddev->private;
7734
7735         if (conf->log || raid5_has_ppl(conf))
7736                 return -EINVAL;
7737         sectors &= ~((sector_t)conf->chunk_sectors - 1);
7738         newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7739         if (mddev->external_size &&
7740             mddev->array_sectors > newsize)
7741                 return -EINVAL;
7742         if (mddev->bitmap) {
7743                 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7744                 if (ret)
7745                         return ret;
7746         }
7747         md_set_array_sectors(mddev, newsize);
7748         if (sectors > mddev->dev_sectors &&
7749             mddev->recovery_cp > mddev->dev_sectors) {
7750                 mddev->recovery_cp = mddev->dev_sectors;
7751                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7752         }
7753         mddev->dev_sectors = sectors;
7754         mddev->resync_max_sectors = sectors;
7755         return 0;
7756 }
7757
7758 static int check_stripe_cache(struct mddev *mddev)
7759 {
7760         /* Can only proceed if there are plenty of stripe_heads.
7761          * We need a minimum of one full stripe,, and for sensible progress
7762          * it is best to have about 4 times that.
7763          * If we require 4 times, then the default 256 4K stripe_heads will
7764          * allow for chunk sizes up to 256K, which is probably OK.
7765          * If the chunk size is greater, user-space should request more
7766          * stripe_heads first.
7767          */
7768         struct r5conf *conf = mddev->private;
7769         if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7770             > conf->min_nr_stripes ||
7771             ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7772             > conf->min_nr_stripes) {
7773                 pr_warn("md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
7774                         mdname(mddev),
7775                         ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7776                          / STRIPE_SIZE)*4);
7777                 return 0;
7778         }
7779         return 1;
7780 }
7781
7782 static int check_reshape(struct mddev *mddev)
7783 {
7784         struct r5conf *conf = mddev->private;
7785
7786         if (conf->log || raid5_has_ppl(conf))
7787                 return -EINVAL;
7788         if (mddev->delta_disks == 0 &&
7789             mddev->new_layout == mddev->layout &&
7790             mddev->new_chunk_sectors == mddev->chunk_sectors)
7791                 return 0; /* nothing to do */
7792         if (has_failed(conf))
7793                 return -EINVAL;
7794         if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7795                 /* We might be able to shrink, but the devices must
7796                  * be made bigger first.
7797                  * For raid6, 4 is the minimum size.
7798                  * Otherwise 2 is the minimum
7799                  */
7800                 int min = 2;
7801                 if (mddev->level == 6)
7802                         min = 4;
7803                 if (mddev->raid_disks + mddev->delta_disks < min)
7804                         return -EINVAL;
7805         }
7806
7807         if (!check_stripe_cache(mddev))
7808                 return -ENOSPC;
7809
7810         if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7811             mddev->delta_disks > 0)
7812                 if (resize_chunks(conf,
7813                                   conf->previous_raid_disks
7814                                   + max(0, mddev->delta_disks),
7815                                   max(mddev->new_chunk_sectors,
7816                                       mddev->chunk_sectors)
7817                             ) < 0)
7818                         return -ENOMEM;
7819
7820         if (conf->previous_raid_disks + mddev->delta_disks <= conf->pool_size)
7821                 return 0; /* never bother to shrink */
7822         return resize_stripes(conf, (conf->previous_raid_disks
7823                                      + mddev->delta_disks));
7824 }
7825
7826 static int raid5_start_reshape(struct mddev *mddev)
7827 {
7828         struct r5conf *conf = mddev->private;
7829         struct md_rdev *rdev;
7830         int spares = 0;
7831         unsigned long flags;
7832
7833         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7834                 return -EBUSY;
7835
7836         if (!check_stripe_cache(mddev))
7837                 return -ENOSPC;
7838
7839         if (has_failed(conf))
7840                 return -EINVAL;
7841
7842         rdev_for_each(rdev, mddev) {
7843                 if (!test_bit(In_sync, &rdev->flags)
7844                     && !test_bit(Faulty, &rdev->flags))
7845                         spares++;
7846         }
7847
7848         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7849                 /* Not enough devices even to make a degraded array
7850                  * of that size
7851                  */
7852                 return -EINVAL;
7853
7854         /* Refuse to reduce size of the array.  Any reductions in
7855          * array size must be through explicit setting of array_size
7856          * attribute.
7857          */
7858         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7859             < mddev->array_sectors) {
7860                 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
7861                         mdname(mddev));
7862                 return -EINVAL;
7863         }
7864
7865         atomic_set(&conf->reshape_stripes, 0);
7866         spin_lock_irq(&conf->device_lock);
7867         write_seqcount_begin(&conf->gen_lock);
7868         conf->previous_raid_disks = conf->raid_disks;
7869         conf->raid_disks += mddev->delta_disks;
7870         conf->prev_chunk_sectors = conf->chunk_sectors;
7871         conf->chunk_sectors = mddev->new_chunk_sectors;
7872         conf->prev_algo = conf->algorithm;
7873         conf->algorithm = mddev->new_layout;
7874         conf->generation++;
7875         /* Code that selects data_offset needs to see the generation update
7876          * if reshape_progress has been set - so a memory barrier needed.
7877          */
7878         smp_mb();
7879         if (mddev->reshape_backwards)
7880                 conf->reshape_progress = raid5_size(mddev, 0, 0);
7881         else
7882                 conf->reshape_progress = 0;
7883         conf->reshape_safe = conf->reshape_progress;
7884         write_seqcount_end(&conf->gen_lock);
7885         spin_unlock_irq(&conf->device_lock);
7886
7887         /* Now make sure any requests that proceeded on the assumption
7888          * the reshape wasn't running - like Discard or Read - have
7889          * completed.
7890          */
7891         mddev_suspend(mddev);
7892         mddev_resume(mddev);
7893
7894         /* Add some new drives, as many as will fit.
7895          * We know there are enough to make the newly sized array work.
7896          * Don't add devices if we are reducing the number of
7897          * devices in the array.  This is because it is not possible
7898          * to correctly record the "partially reconstructed" state of
7899          * such devices during the reshape and confusion could result.
7900          */
7901         if (mddev->delta_disks >= 0) {
7902                 rdev_for_each(rdev, mddev)
7903                         if (rdev->raid_disk < 0 &&
7904                             !test_bit(Faulty, &rdev->flags)) {
7905                                 if (raid5_add_disk(mddev, rdev) == 0) {
7906                                         if (rdev->raid_disk
7907                                             >= conf->previous_raid_disks)
7908                                                 set_bit(In_sync, &rdev->flags);
7909                                         else
7910                                                 rdev->recovery_offset = 0;
7911
7912                                         if (sysfs_link_rdev(mddev, rdev))
7913                                                 /* Failure here is OK */;
7914                                 }
7915                         } else if (rdev->raid_disk >= conf->previous_raid_disks
7916                                    && !test_bit(Faulty, &rdev->flags)) {
7917                                 /* This is a spare that was manually added */
7918                                 set_bit(In_sync, &rdev->flags);
7919                         }
7920
7921                 /* When a reshape changes the number of devices,
7922                  * ->degraded is measured against the larger of the
7923                  * pre and post number of devices.
7924                  */
7925                 spin_lock_irqsave(&conf->device_lock, flags);
7926                 mddev->degraded = raid5_calc_degraded(conf);
7927                 spin_unlock_irqrestore(&conf->device_lock, flags);
7928         }
7929         mddev->raid_disks = conf->raid_disks;
7930         mddev->reshape_position = conf->reshape_progress;
7931         set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
7932
7933         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7934         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7935         clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7936         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7937         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7938         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7939                                                 "reshape");
7940         if (!mddev->sync_thread) {
7941                 mddev->recovery = 0;
7942                 spin_lock_irq(&conf->device_lock);
7943                 write_seqcount_begin(&conf->gen_lock);
7944                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7945                 mddev->new_chunk_sectors =
7946                         conf->chunk_sectors = conf->prev_chunk_sectors;
7947                 mddev->new_layout = conf->algorithm = conf->prev_algo;
7948                 rdev_for_each(rdev, mddev)
7949                         rdev->new_data_offset = rdev->data_offset;
7950                 smp_wmb();
7951                 conf->generation --;
7952                 conf->reshape_progress = MaxSector;
7953                 mddev->reshape_position = MaxSector;
7954                 write_seqcount_end(&conf->gen_lock);
7955                 spin_unlock_irq(&conf->device_lock);
7956                 return -EAGAIN;
7957         }
7958         conf->reshape_checkpoint = jiffies;
7959         md_wakeup_thread(mddev->sync_thread);
7960         md_new_event(mddev);
7961         return 0;
7962 }
7963
7964 /* This is called from the reshape thread and should make any
7965  * changes needed in 'conf'
7966  */
7967 static void end_reshape(struct r5conf *conf)
7968 {
7969
7970         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7971
7972                 spin_lock_irq(&conf->device_lock);
7973                 conf->previous_raid_disks = conf->raid_disks;
7974                 md_finish_reshape(conf->mddev);
7975                 smp_wmb();
7976                 conf->reshape_progress = MaxSector;
7977                 conf->mddev->reshape_position = MaxSector;
7978                 spin_unlock_irq(&conf->device_lock);
7979                 wake_up(&conf->wait_for_overlap);
7980
7981                 /* read-ahead size must cover two whole stripes, which is
7982                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7983                  */
7984                 if (conf->mddev->queue) {
7985                         int data_disks = conf->raid_disks - conf->max_degraded;
7986                         int stripe = data_disks * ((conf->chunk_sectors << 9)
7987                                                    / PAGE_SIZE);
7988                         if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
7989                                 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
7990                 }
7991         }
7992 }
7993
7994 /* This is called from the raid5d thread with mddev_lock held.
7995  * It makes config changes to the device.
7996  */
7997 static void raid5_finish_reshape(struct mddev *mddev)
7998 {
7999         struct r5conf *conf = mddev->private;
8000
8001         if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
8002
8003                 if (mddev->delta_disks > 0) {
8004                         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
8005                         if (mddev->queue) {
8006                                 set_capacity(mddev->gendisk, mddev->array_sectors);
8007                                 revalidate_disk(mddev->gendisk);
8008                         }
8009                 } else {
8010                         int d;
8011                         spin_lock_irq(&conf->device_lock);
8012                         mddev->degraded = raid5_calc_degraded(conf);
8013                         spin_unlock_irq(&conf->device_lock);
8014                         for (d = conf->raid_disks ;
8015                              d < conf->raid_disks - mddev->delta_disks;
8016                              d++) {
8017                                 struct md_rdev *rdev = conf->disks[d].rdev;
8018                                 if (rdev)
8019                                         clear_bit(In_sync, &rdev->flags);
8020                                 rdev = conf->disks[d].replacement;
8021                                 if (rdev)
8022                                         clear_bit(In_sync, &rdev->flags);
8023                         }
8024                 }
8025                 mddev->layout = conf->algorithm;
8026                 mddev->chunk_sectors = conf->chunk_sectors;
8027                 mddev->reshape_position = MaxSector;
8028                 mddev->delta_disks = 0;
8029                 mddev->reshape_backwards = 0;
8030         }
8031 }
8032
8033 static void raid5_quiesce(struct mddev *mddev, int state)
8034 {
8035         struct r5conf *conf = mddev->private;
8036
8037         switch(state) {
8038         case 2: /* resume for a suspend */
8039                 wake_up(&conf->wait_for_overlap);
8040                 break;
8041
8042         case 1: /* stop all writes */
8043                 lock_all_device_hash_locks_irq(conf);
8044                 /* '2' tells resync/reshape to pause so that all
8045                  * active stripes can drain
8046                  */
8047                 r5c_flush_cache(conf, INT_MAX);
8048                 conf->quiesce = 2;
8049                 wait_event_cmd(conf->wait_for_quiescent,
8050                                     atomic_read(&conf->active_stripes) == 0 &&
8051                                     atomic_read(&conf->active_aligned_reads) == 0,
8052                                     unlock_all_device_hash_locks_irq(conf),
8053                                     lock_all_device_hash_locks_irq(conf));
8054                 conf->quiesce = 1;
8055                 unlock_all_device_hash_locks_irq(conf);
8056                 /* allow reshape to continue */
8057                 wake_up(&conf->wait_for_overlap);
8058                 break;
8059
8060         case 0: /* re-enable writes */
8061                 lock_all_device_hash_locks_irq(conf);
8062                 conf->quiesce = 0;
8063                 wake_up(&conf->wait_for_quiescent);
8064                 wake_up(&conf->wait_for_overlap);
8065                 unlock_all_device_hash_locks_irq(conf);
8066                 break;
8067         }
8068         r5l_quiesce(conf->log, state);
8069 }
8070
8071 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
8072 {
8073         struct r0conf *raid0_conf = mddev->private;
8074         sector_t sectors;
8075
8076         /* for raid0 takeover only one zone is supported */
8077         if (raid0_conf->nr_strip_zones > 1) {
8078                 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8079                         mdname(mddev));
8080                 return ERR_PTR(-EINVAL);
8081         }
8082
8083         sectors = raid0_conf->strip_zone[0].zone_end;
8084         sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
8085         mddev->dev_sectors = sectors;
8086         mddev->new_level = level;
8087         mddev->new_layout = ALGORITHM_PARITY_N;
8088         mddev->new_chunk_sectors = mddev->chunk_sectors;
8089         mddev->raid_disks += 1;
8090         mddev->delta_disks = 1;
8091         /* make sure it will be not marked as dirty */
8092         mddev->recovery_cp = MaxSector;
8093
8094         return setup_conf(mddev);
8095 }
8096
8097 static void *raid5_takeover_raid1(struct mddev *mddev)
8098 {
8099         int chunksect;
8100         void *ret;
8101
8102         if (mddev->raid_disks != 2 ||
8103             mddev->degraded > 1)
8104                 return ERR_PTR(-EINVAL);
8105
8106         /* Should check if there are write-behind devices? */
8107
8108         chunksect = 64*2; /* 64K by default */
8109
8110         /* The array must be an exact multiple of chunksize */
8111         while (chunksect && (mddev->array_sectors & (chunksect-1)))
8112                 chunksect >>= 1;
8113
8114         if ((chunksect<<9) < STRIPE_SIZE)
8115                 /* array size does not allow a suitable chunk size */
8116                 return ERR_PTR(-EINVAL);
8117
8118         mddev->new_level = 5;
8119         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
8120         mddev->new_chunk_sectors = chunksect;
8121
8122         ret = setup_conf(mddev);
8123         if (!IS_ERR(ret))
8124                 mddev_clear_unsupported_flags(mddev,
8125                         UNSUPPORTED_MDDEV_FLAGS);
8126         return ret;
8127 }
8128
8129 static void *raid5_takeover_raid6(struct mddev *mddev)
8130 {
8131         int new_layout;
8132
8133         switch (mddev->layout) {
8134         case ALGORITHM_LEFT_ASYMMETRIC_6:
8135                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
8136                 break;
8137         case ALGORITHM_RIGHT_ASYMMETRIC_6:
8138                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
8139                 break;
8140         case ALGORITHM_LEFT_SYMMETRIC_6:
8141                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
8142                 break;
8143         case ALGORITHM_RIGHT_SYMMETRIC_6:
8144                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
8145                 break;
8146         case ALGORITHM_PARITY_0_6:
8147                 new_layout = ALGORITHM_PARITY_0;
8148                 break;
8149         case ALGORITHM_PARITY_N:
8150                 new_layout = ALGORITHM_PARITY_N;
8151                 break;
8152         default:
8153                 return ERR_PTR(-EINVAL);
8154         }
8155         mddev->new_level = 5;
8156         mddev->new_layout = new_layout;
8157         mddev->delta_disks = -1;
8158         mddev->raid_disks -= 1;
8159         return setup_conf(mddev);
8160 }
8161
8162 static int raid5_check_reshape(struct mddev *mddev)
8163 {
8164         /* For a 2-drive array, the layout and chunk size can be changed
8165          * immediately as not restriping is needed.
8166          * For larger arrays we record the new value - after validation
8167          * to be used by a reshape pass.
8168          */
8169         struct r5conf *conf = mddev->private;
8170         int new_chunk = mddev->new_chunk_sectors;
8171
8172         if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
8173                 return -EINVAL;
8174         if (new_chunk > 0) {
8175                 if (!is_power_of_2(new_chunk))
8176                         return -EINVAL;
8177                 if (new_chunk < (PAGE_SIZE>>9))
8178                         return -EINVAL;
8179                 if (mddev->array_sectors & (new_chunk-1))
8180                         /* not factor of array size */
8181                         return -EINVAL;
8182         }
8183
8184         /* They look valid */
8185
8186         if (mddev->raid_disks == 2) {
8187                 /* can make the change immediately */
8188                 if (mddev->new_layout >= 0) {
8189                         conf->algorithm = mddev->new_layout;
8190                         mddev->layout = mddev->new_layout;
8191                 }
8192                 if (new_chunk > 0) {
8193                         conf->chunk_sectors = new_chunk ;
8194                         mddev->chunk_sectors = new_chunk;
8195                 }
8196                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8197                 md_wakeup_thread(mddev->thread);
8198         }
8199         return check_reshape(mddev);
8200 }
8201
8202 static int raid6_check_reshape(struct mddev *mddev)
8203 {
8204         int new_chunk = mddev->new_chunk_sectors;
8205
8206         if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
8207                 return -EINVAL;
8208         if (new_chunk > 0) {
8209                 if (!is_power_of_2(new_chunk))
8210                         return -EINVAL;
8211                 if (new_chunk < (PAGE_SIZE >> 9))
8212                         return -EINVAL;
8213                 if (mddev->array_sectors & (new_chunk-1))
8214                         /* not factor of array size */
8215                         return -EINVAL;
8216         }
8217
8218         /* They look valid */
8219         return check_reshape(mddev);
8220 }
8221
8222 static void *raid5_takeover(struct mddev *mddev)
8223 {
8224         /* raid5 can take over:
8225          *  raid0 - if there is only one strip zone - make it a raid4 layout
8226          *  raid1 - if there are two drives.  We need to know the chunk size
8227          *  raid4 - trivial - just use a raid4 layout.
8228          *  raid6 - Providing it is a *_6 layout
8229          */
8230         if (mddev->level == 0)
8231                 return raid45_takeover_raid0(mddev, 5);
8232         if (mddev->level == 1)
8233                 return raid5_takeover_raid1(mddev);
8234         if (mddev->level == 4) {
8235                 mddev->new_layout = ALGORITHM_PARITY_N;
8236                 mddev->new_level = 5;
8237                 return setup_conf(mddev);
8238         }
8239         if (mddev->level == 6)
8240                 return raid5_takeover_raid6(mddev);
8241
8242         return ERR_PTR(-EINVAL);
8243 }
8244
8245 static void *raid4_takeover(struct mddev *mddev)
8246 {
8247         /* raid4 can take over:
8248          *  raid0 - if there is only one strip zone
8249          *  raid5 - if layout is right
8250          */
8251         if (mddev->level == 0)
8252                 return raid45_takeover_raid0(mddev, 4);
8253         if (mddev->level == 5 &&
8254             mddev->layout == ALGORITHM_PARITY_N) {
8255                 mddev->new_layout = 0;
8256                 mddev->new_level = 4;
8257                 return setup_conf(mddev);
8258         }
8259         return ERR_PTR(-EINVAL);
8260 }
8261
8262 static struct md_personality raid5_personality;
8263
8264 static void *raid6_takeover(struct mddev *mddev)
8265 {
8266         /* Currently can only take over a raid5.  We map the
8267          * personality to an equivalent raid6 personality
8268          * with the Q block at the end.
8269          */
8270         int new_layout;
8271
8272         if (mddev->pers != &raid5_personality)
8273                 return ERR_PTR(-EINVAL);
8274         if (mddev->degraded > 1)
8275                 return ERR_PTR(-EINVAL);
8276         if (mddev->raid_disks > 253)
8277                 return ERR_PTR(-EINVAL);
8278         if (mddev->raid_disks < 3)
8279                 return ERR_PTR(-EINVAL);
8280
8281         switch (mddev->layout) {
8282         case ALGORITHM_LEFT_ASYMMETRIC:
8283                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
8284                 break;
8285         case ALGORITHM_RIGHT_ASYMMETRIC:
8286                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
8287                 break;
8288         case ALGORITHM_LEFT_SYMMETRIC:
8289                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
8290                 break;
8291         case ALGORITHM_RIGHT_SYMMETRIC:
8292                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
8293                 break;
8294         case ALGORITHM_PARITY_0:
8295                 new_layout = ALGORITHM_PARITY_0_6;
8296                 break;
8297         case ALGORITHM_PARITY_N:
8298                 new_layout = ALGORITHM_PARITY_N;
8299                 break;
8300         default:
8301                 return ERR_PTR(-EINVAL);
8302         }
8303         mddev->new_level = 6;
8304         mddev->new_layout = new_layout;
8305         mddev->delta_disks = 1;
8306         mddev->raid_disks += 1;
8307         return setup_conf(mddev);
8308 }
8309
8310 static int raid5_change_consistency_policy(struct mddev *mddev, const char *buf)
8311 {
8312         struct r5conf *conf;
8313         int err;
8314
8315         err = mddev_lock(mddev);
8316         if (err)
8317                 return err;
8318         conf = mddev->private;
8319         if (!conf) {
8320                 mddev_unlock(mddev);
8321                 return -ENODEV;
8322         }
8323
8324         if (strncmp(buf, "ppl", 3) == 0) {
8325                 /* ppl only works with RAID 5 */
8326                 if (!raid5_has_ppl(conf) && conf->level == 5) {
8327                         err = log_init(conf, NULL, true);
8328                         if (!err) {
8329                                 err = resize_stripes(conf, conf->pool_size);
8330                                 if (err)
8331                                         log_exit(conf);
8332                         }
8333                 } else
8334                         err = -EINVAL;
8335         } else if (strncmp(buf, "resync", 6) == 0) {
8336                 if (raid5_has_ppl(conf)) {
8337                         mddev_suspend(mddev);
8338                         log_exit(conf);
8339                         mddev_resume(mddev);
8340                         err = resize_stripes(conf, conf->pool_size);
8341                 } else if (test_bit(MD_HAS_JOURNAL, &conf->mddev->flags) &&
8342                            r5l_log_disk_error(conf)) {
8343                         bool journal_dev_exists = false;
8344                         struct md_rdev *rdev;
8345
8346                         rdev_for_each(rdev, mddev)
8347                                 if (test_bit(Journal, &rdev->flags)) {
8348                                         journal_dev_exists = true;
8349                                         break;
8350                                 }
8351
8352                         if (!journal_dev_exists) {
8353                                 mddev_suspend(mddev);
8354                                 clear_bit(MD_HAS_JOURNAL, &mddev->flags);
8355                                 mddev_resume(mddev);
8356                         } else  /* need remove journal device first */
8357                                 err = -EBUSY;
8358                 } else
8359                         err = -EINVAL;
8360         } else {
8361                 err = -EINVAL;
8362         }
8363
8364         if (!err)
8365                 md_update_sb(mddev, 1);
8366
8367         mddev_unlock(mddev);
8368
8369         return err;
8370 }
8371
8372 static struct md_personality raid6_personality =
8373 {
8374         .name           = "raid6",
8375         .level          = 6,
8376         .owner          = THIS_MODULE,
8377         .make_request   = raid5_make_request,
8378         .run            = raid5_run,
8379         .free           = raid5_free,
8380         .status         = raid5_status,
8381         .error_handler  = raid5_error,
8382         .hot_add_disk   = raid5_add_disk,
8383         .hot_remove_disk= raid5_remove_disk,
8384         .spare_active   = raid5_spare_active,
8385         .sync_request   = raid5_sync_request,
8386         .resize         = raid5_resize,
8387         .size           = raid5_size,
8388         .check_reshape  = raid6_check_reshape,
8389         .start_reshape  = raid5_start_reshape,
8390         .finish_reshape = raid5_finish_reshape,
8391         .quiesce        = raid5_quiesce,
8392         .takeover       = raid6_takeover,
8393         .congested      = raid5_congested,
8394         .change_consistency_policy = raid5_change_consistency_policy,
8395 };
8396 static struct md_personality raid5_personality =
8397 {
8398         .name           = "raid5",
8399         .level          = 5,
8400         .owner          = THIS_MODULE,
8401         .make_request   = raid5_make_request,
8402         .run            = raid5_run,
8403         .free           = raid5_free,
8404         .status         = raid5_status,
8405         .error_handler  = raid5_error,
8406         .hot_add_disk   = raid5_add_disk,
8407         .hot_remove_disk= raid5_remove_disk,
8408         .spare_active   = raid5_spare_active,
8409         .sync_request   = raid5_sync_request,
8410         .resize         = raid5_resize,
8411         .size           = raid5_size,
8412         .check_reshape  = raid5_check_reshape,
8413         .start_reshape  = raid5_start_reshape,
8414         .finish_reshape = raid5_finish_reshape,
8415         .quiesce        = raid5_quiesce,
8416         .takeover       = raid5_takeover,
8417         .congested      = raid5_congested,
8418         .change_consistency_policy = raid5_change_consistency_policy,
8419 };
8420
8421 static struct md_personality raid4_personality =
8422 {
8423         .name           = "raid4",
8424         .level          = 4,
8425         .owner          = THIS_MODULE,
8426         .make_request   = raid5_make_request,
8427         .run            = raid5_run,
8428         .free           = raid5_free,
8429         .status         = raid5_status,
8430         .error_handler  = raid5_error,
8431         .hot_add_disk   = raid5_add_disk,
8432         .hot_remove_disk= raid5_remove_disk,
8433         .spare_active   = raid5_spare_active,
8434         .sync_request   = raid5_sync_request,
8435         .resize         = raid5_resize,
8436         .size           = raid5_size,
8437         .check_reshape  = raid5_check_reshape,
8438         .start_reshape  = raid5_start_reshape,
8439         .finish_reshape = raid5_finish_reshape,
8440         .quiesce        = raid5_quiesce,
8441         .takeover       = raid4_takeover,
8442         .congested      = raid5_congested,
8443         .change_consistency_policy = raid5_change_consistency_policy,
8444 };
8445
8446 static int __init raid5_init(void)
8447 {
8448         int ret;
8449
8450         raid5_wq = alloc_workqueue("raid5wq",
8451                 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
8452         if (!raid5_wq)
8453                 return -ENOMEM;
8454
8455         ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
8456                                       "md/raid5:prepare",
8457                                       raid456_cpu_up_prepare,
8458                                       raid456_cpu_dead);
8459         if (ret) {
8460                 destroy_workqueue(raid5_wq);
8461                 return ret;
8462         }
8463         register_md_personality(&raid6_personality);
8464         register_md_personality(&raid5_personality);
8465         register_md_personality(&raid4_personality);
8466         return 0;
8467 }
8468
8469 static void raid5_exit(void)
8470 {
8471         unregister_md_personality(&raid6_personality);
8472         unregister_md_personality(&raid5_personality);
8473         unregister_md_personality(&raid4_personality);
8474         cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
8475         destroy_workqueue(raid5_wq);
8476 }
8477
8478 module_init(raid5_init);
8479 module_exit(raid5_exit);
8480 MODULE_LICENSE("GPL");
8481 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8482 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8483 MODULE_ALIAS("md-raid5");
8484 MODULE_ALIAS("md-raid4");
8485 MODULE_ALIAS("md-level-5");
8486 MODULE_ALIAS("md-level-4");
8487 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8488 MODULE_ALIAS("md-raid6");
8489 MODULE_ALIAS("md-level-6");
8490
8491 /* This used to be two separate modules, they were: */
8492 MODULE_ALIAS("raid5");
8493 MODULE_ALIAS("raid6");